run_id: g1-piston-starvla-oft-v2
run_root_dir: /home/bir/gr00t_checkpoints
seed: 42
wandb_entity: your_wandb_entity
wandb_project: starVLA_unitree_g1_piston
is_debug: false
version_id: '0.21'
framework:
  name: QwenOFT
  qwenvl:
    base_vlm: ./playground/Pretrained_models/Qwen3-VL-2B-Instruct
    attn_implementation: sdpa
    enable_gradient_checkpointing: true
  action_model:
    action_model_type: MLP
    action_dim: 30
    action_hidden_dim: 2048
    future_action_window_size: 29
    past_action_window_size: 0
    state_dim: 0
    action_horizon: 30
datasets:
  vla_data:
    dataset_py: lerobot_datasets
    include_state: false
    data_root_dir: /home/bir/datasets
    data_mix: unitree_g1_piston_pnp
    lerobot_version: v2.0
    action_type: abs_qpos
    action_mode: abs
    sequential_step_sampling: false
    per_device_batch_size: 8
    load_all_data_for_training: true
    obs_image_size:
    - 224
    - 224
    delete_pause_frame: false
    video_backend: torchvision_av
    num_workers: 4
trainer:
  max_train_steps: 10000
  num_warmup_steps: 50
  save_interval: 1000
  eval_interval: 500
  learning_rate:
    base: 2.0e-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
  weight_decay: 0.0
  logging_frequency: 10
  gradient_clipping: 1.0
  gradient_accumulation_steps: 4
  gradient_checkpointing: true
  optimizer:
    name: paged_adamw_8bit
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  save_format: pt
config_yaml: examples/realRobots/UnitreeG1_Piston/step2_training/train_files/starvla_qwenoft_g1_piston_v2.yaml
output_dir: /home/bir/gr00t_checkpoints/g1-piston-starvla-oft-v2