run_id: subtask_m2w_libero_implicit_jepa_query_8
run_root_dir: results
seed: 42
wandb_entity: livion
wandb_project: m2w-vla
is_debug: false
framework:
  name: QwenSubtaskM2W_phs
  qwenvl:
    base_vlm: ./playground/Pretrained_models/Qwen3-VL-4B-Instruct
    attn_implementation: sdpa
    vl_hidden_dim: 2048
  reasoning:
    cot_reasoning_mode: implicit_query
    latent_cot_num_queries: 8
    qwen_action_context_mode: exclude_cot_prompt
    lambda_cot: 0.1
    generate_cot_at_inference: false
    inference_cot_max_new_tokens: 96
    inference_cot_use_cache: true
    inference_empty_cache: false
  vjepa2:
    hub_repo: facebookresearch/vjepa2:main
    hub_model: vjepa2_1_vit_large_384
    hub_source: github
    num_frames: 8
    image_size: 224
  jepa_predictor:
    prediction_view: wrist
    dropout: 0.0
    use_jepa_loss: true
    lambda_jepa: 0.2
    jepa_predictor_depth: 4
    jepa_predictor_num_heads: 8
    jepa_predictor_bf16_autocast: true
    jepa_predictor_activation_checkpointing: true
  policy_context:
    use_state: false
    wrist_action_context_tokens: 32
    wrist_action_context_heads: 8
    wrist_action_context_layers: 1
    detach_wrist_action_context: true
  action_model:
    action_model_type: DiT-B
    gradient_checkpointing: true
    action_hidden_dim: 1024
    hidden_size: 1024
    add_pos_embed: true
    max_seq_len: 1024
    action_dim: 7
    state_dim: null
    action_horizon: 8
    past_action_window_size: 0
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_timestep_buckets: 1000
    num_inference_timesteps: 4
    num_target_vision_tokens: 32
    diffusion_model_cfg:
      cross_attention_dim: 2560
      dropout: 0.2
      final_dropout: true
      interleave_self_attention: true
      norm_type: ada_norm
      num_layers: 16
      output_dim: 1024
      positional_embeddings: null
datasets:
  vla_data:
    dataset_py: subtask_m2w_datasets_phs
    data_root_dir: playground/Datasets/LEROBOT_LIBERO_DATA
    data_mix: libero_all
    action_type: delta_qpos
    per_device_batch_size: 16
    load_all_data_for_training: true
    video_backend: torchvision_av
    jepa_prediction_view: wrist
    vlm_include_wrist_view: true
    wrist_history_frames: 8
    future_wrist_frames: 8
    future_wrist_k: 8
    subtask_label_dir: playground/Datasets/subtask_m2w_labels_libero_qwen3vl_full_softqa_v31_compact_final_v16_20260630
    CoT_prompt: 'Your task is: {instruction}

      Using the current main-view and wrist-view images, identify the robot''s current
      manipulation step.

      Output exactly:

      Subtask: ...

      Reasoning: ...

      Wrist: ...

      Do not add extra text.'
trainer:
  max_train_steps: 30000
  num_warmup_steps: 2500
  is_resume: false
  pretrained_checkpoint: null
  reload_modules: null
  resume_optimizer_state: true
  train_qwen_vl: true
  save_interval: 10000
  logging_frequency: 100
  repeated_diffusion_steps: 8
  learning_rate:
    base: 2.5e-05
    jepa_predictor: 5.0e-05
    wrist_context_adapter: 5.0e-05
    action_model: 0.0001
    vlm: 1.0e-05
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
  gradient_clipping: 1.0
  gradient_accumulation_steps: 1
  optimizer:
    name: AdamW
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
output_dir: results/subtask_m2w_libero_implicit_jepa_query_8