run_id: LIBERO
run_root_dir: checkpoints
seed: 42
trackers:
- json
is_debug: false
framework:
  name: VLA_JEPA
  qwenvl:
    base_vlm: /home/dataset-local/models/Qwen3-VL-2B-Instruct
    attn_implementation: flash_attention_2
    vl_hidden_dim: 2048
  action_model:
    action_model_type: DiT-B
    action_hidden_dim: 1024
    hidden_size: 1024
    add_pos_embed: true
    max_seq_len: 1024
    action_dim: 7
    state_dim: 8
    future_action_window_size: 6
    action_horizon: 7
    past_action_window_size: 0
    repeated_diffusion_steps: 8
    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: 2048
      dropout: 0.2
      final_dropout: true
      interleave_self_attention: true
      norm_type: ada_norm
      num_layers: 16
      output_dim: 1024
      positional_embeddings: null
  vj2_model:
    base_encoder: /home/dataset-local/models/vjepa2-vitl-fpc64-256
    depth: 12
    num_heads: 8
    special_action_token: <|action_{}|>
    num_action_tokens_per_timestep: 8
    embodied_action_token: <|embodied_action|>
    num_embodied_action_tokens_per_instruction: 32
    num_frames: 8
  reduce_in_full_precision: true
datasets:
  vla_data:
    dataset_py: lerobot_datasets
    data_root_dir: /home/dataset-local/datasets/LeRobot/LEROBOT_LIBERO_DATA
    data_mix: libero_all
    action_type: delta_qpos
    CoT_prompt: Your task is {instruction}. Infer the temporal dynamics from frames
      {actions} and produce the corresponding policy actions {e_actions}.
    resolution_size: 224
    per_device_batch_size: 32
    video_resolution_size: 256
    load_all_data_for_training: true
    with_state: true
trainer:
  epochs: 100
  max_train_steps: 30000
  num_warmup_steps: 5000
  save_interval: 10000
  eval_interval: 100
  learning_rate:
    base: 3.0e-05
    qwen_vl_interface: 1.0e-05
    action_model: 0.0001
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
  freeze_modules: ''
  loss_scale:
    vla: 1.0
    vlm: 0.1
  max_grad_norm: 1.0
  warmup_ratio: 0.1
  weight_decay: 0.0
  logging_frequency: 10
  gradient_clipping: 1.0
  gradient_accumulation_steps: 1
  pretrained_checkpoint: /home/dataset-local/VLA_JEPA/checkpoints/pretrain/VLA-JEPA-pretrain.pt
  optimizer:
    name: AdamW
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  is_resume: false
  resume_epoch: null
  resume_step: null
  enable_gradient_checkpointing: true
  enable_mixed_precision_training: true
output_dir: checkpoints/LIBERO