checkpoint:
  local:
    keep_last_n: 1
  save_best_model: true
  save_pt_file: false
  sync:
    enabled: false
    keep_last_n: 0
    repo_id: null
datasets:
  vla_data:
    data_mix: demon_attack_train__bridge
    dataset_py: lerobot_datasets
    eval_data_mix: demon_attack_train__bridge__val
    include_state: true
    obs_image_size: null
    per_device_batch_size: 32
framework:
  action_model:
    action_dim: 7
    action_env_dim: 6
    action_hidden_dim: 2560
    action_horizon: 1
    action_model_type: MLP
    loss_type: discrete_ce
    state_dim: 7
  name: QwenOFT
  qwenvl:
    attn_implementation: flash_attention_2
    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
    enable_gradient_checkpointing: true
output_dir: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2_small
rl_games:
  env_eval:
    demon_attack:
      noop_max: 30
    distributed_mode: none
    enabled: true
    fixed_episode_seeds: true
    frameskip: 4
    image_size: 224
    latency:
      prompt_map_path: /workspace/starVLA/data/demon_attack_fix_latency_2_small/demon_attack_train__bridge/latency_prompt_map.json
    latency_seed_stride: 0
    mid_train:
      enabled: true
      interval_steps: 100
      latencies:
      - 2
      max_steps_per_episode: 3600
      num_episodes: 5
    seed: 42
    task_description: 'You are playing Demon Attack from a single game image. Choose
      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
    task_seed_stride: 0
    vectorized:
      enabled: false
  model_alias: openvla
  task: demon_attack
run_id: openvla_demon_attack_fix_latency_2_small
run_root_dir: /workspace/starVLA/results/Checkpoints
seed: 42
trainer:
  distributed_backend: none
  eval_interval: 100
  eval_num_batches: 20
  freeze_modules: ''
  gradient_accumulation_steps: 16
  gradient_clipping: 1.0
  is_resume: true
  learning_rate:
    action_model: 0.0001
    base: 2.0e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 10
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 2000
  num_warmup_steps: 100
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    fused: false
    weight_decay: 1.0e-08
  pretrained_checkpoint: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_600_state
  resume_step: 600
  save_interval: 100
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
wandb_entity: dongqianyu99-zhejiang-university
wandb_project: starVLA_rl_games