datasets:
  vla_data:
    data_mix: so101_pickorange
    data_root_dir: /root/autodl-tmp/datasets
    dataset_py: lerobot_datasets
    lerobot_version: v2.0
    per_device_batch_size: 8
    video_backend: torchvision_av
framework:
  action_model:
    action_dim: 6
    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
    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
    repeated_diffusion_steps: 8
    state_dim: 6
  name: QwenGR00T
  qwenvl:
    attn_implementation: flash_attention_2
    base_vlm: /home/david/.cache/huggingface/hub/models--Qwen--Qwen3-VL-4B-Instruct/snapshots/ebb281ec70b05090aa6165b016eac8ec08e71b17
output_dir: /root/autodl-tmp/starvla-outputs/so101_pickorange_qwengr00t
run_id: so101_pickorange_qwengr00t
run_root_dir: /root/autodl-tmp/starvla-outputs
seed: 42
trainer:
  eval_interval: 100000
  freeze_modules: qwen_vl_interface
  gradient_clipping: 1.0
  keep_last_checkpoints: 2
  learning_rate:
    action_model: 0.0001
    base: 2.5e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 10
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 30000
  num_warmup_steps: 1500
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  save_interval: 500
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
wandb_entity: ''
wandb_project: starVLA_SO101