datasets:
  vla_data:
    data_mix: robotwin
    data_root_dir: /inspire/ssd/project/embodied-basic-model/zhangjianing-253108140206/DATASET/robotwin_lerobot
    dataset_py: lerobot_datasets
    image_size:
    - 224
    - 224
    per_device_batch_size: 8
    video_backend: torchvision_av
framework:
  action_model:
    action_dim: 14
    action_hidden_dim: 2560
    action_model_type: DiT-B
    future_action_window_size: 15
    past_action_window_size: 0
  name: QwenOFT
  qwenvl:
    base_vlm: /inspire/ssd/project/embodied-basic-model/zhangjianing-253108140206/DATASET/model/spiritv1.5
output_dir: /inspire/ssd/project/embodied-basic-model/zhangjianing-253108140206/experiment/spirit_vla/starvla-vla/results/124_robotwin_spirit
run_id: 124_robotwin_spirit
run_root_dir: /inspire/ssd/project/embodied-basic-model/zhangjianing-253108140206/experiment/spirit_vla/starvla-vla/results
seed: 42
trainer:
  eval_interval: 1000
  freeze_modules: true
  gradient_accumulation_steps: 1
  gradient_clipping: 1.0
  is_resume: false
  learning_rate:
    action_model: 0.0001
    base: 1.0e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 100
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 30000
  num_warmup_steps: 5000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  save_interval: 10000
  scheduler_specific_kwargs:
    min_lr: 5.0e-07
wandb_entity: 1732949190-tongji-university
wandb_project: spirit