datasets:
  vla_data:
    data_mix: delta_state
    data_root_dir: /mnt/xlab-nas-2/vla_dataset
    dataset_py: lerobot_datasets
    image_size:
    - 224
    - 224
    num_workers: 6
    per_device_batch_size: 16
    sequential_step_sampling: false
framework:
  action_model:
    action_dim: 14
    action_horizon: 16
    action_model_type: DiT-B
    add_pos_embed: true
    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
    future_action_window_size: 15
    hidden_size: 1024
    max_seq_len: 1024
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_inference_timesteps: 4
    num_target_vision_tokens: 32
    num_timestep_buckets: 1000
    past_action_window_size: 0
    state_dim: 14
  name: QwenJAT
  qwenvl:
    base_vlm: ./checkpoints/Qwen3-VL-4B-Instruct-Action
output_dir: /mnt/workspace/lintong.lt/output/vla_pretrain/0323_pretrain_Qwen3VL4BJAT_bs2048
run_id: 0323_pretrain_Qwen3VL4BJAT_bs2048
run_root_dir: /mnt/workspace/lintong.lt/output/vla_pretrain
seed: 42
trainer:
  eval_interval: 2000
  freeze_modules: null
  gradient_accumulation_steps: 2
  gradient_clipping: 1.0
  is_resume: false
  learning_rate:
    action_model: 0.0001
    base: 3.0e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 100
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 200000
  num_warmup_steps: 2000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  repeated_diffusion_steps: 4
  save_interval: 2000
  scheduler_specific_kwargs:
    min_lr: 5.0e-07
wandb_entity: your_wandb_entity
wandb_project: llavavla