datasets:
  vla_data:
    data_mix: real_franka_v4
    data_root_dir: /gpfs/wangzixuan/Octopus/workspace/data/lerobot/v4
    dataset_py: lerobot_datasets
    per_device_batch_size: 16
    sequential_step_sampling: false
    video_backend: torchvision_av
framework:
  action_model:
    action_dim: 7
    action_hidden_dim: 2560
    action_model_type: DiT-B
    future_action_window_size: 15
    past_action_window_size: 0
  name: QwenOFT
  obs_image_size: null
  qwenvl:
    attn_implementation: flash_attention_2
    base_vlm: /gpfs/wangzixuan/Octopus/workspace/MODEL/Qwen3-VL-4B-Instruct
output_dir: /gpfs/wangzixuan/Octopus/workspace/starVLA/results/Checkpoints/0622_franka_v4_qwen3OFT_sandwich
run_id: 0622_franka_v4_qwen3OFT_sandwich
run_root_dir: /gpfs/wangzixuan/Octopus/workspace/starVLA/results/Checkpoints
seed: 42
trainer:
  eval_interval: 1000
  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: 100000
  num_warmup_steps: 5000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  pretrained_checkpoint: ''
  save_interval: 10000
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
wandb_entity: zwanggk
wandb_project: roboclaw