run_id: 0615_franka_v3_dual_finetune_qwen3OFT
run_root_dir: /gpfs/wangzixuan/Octopus/workspace/starVLA/results/Checkpoints
seed: 42
trackers:
- jsonl
- wandb
wandb_entity: zwanggk
wandb_project: franka_dual
is_debug: false
framework:
  name: QwenOFT
  qwenvl:
    base_vlm: /gpfs/wangzixuan/Octopus/workspace/MODEL/Qwen3-VL-4B-Instruct
    attn_implementation: flash_attention_2
  action_model:
    action_model_type: DiT-B
    action_dim: 14
    action_hidden_dim: 2560
    future_action_window_size: 15
    past_action_window_size: 0
    state_dim: 56
  obs_image_size: null
datasets:
  vla_data:
    dataset_py: lerobot_datasets
    data_root_dir: /gpfs/wangzixuan/Octopus/workspace/data/lerobot/v3
    data_mix: real_franka_dual_v3
    image_size:
    - 224
    - 224
    per_device_batch_size: 16
    sequential_step_sampling: false
    video_backend: torchvision_av
trainer:
  pretrained_checkpoint: ''
  reload_modules: qwen_vl_interface
  enable_gradient_checkpointing: true
  enable_mixed_precision_training: true
  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:
    name: AdamW
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  save_interval: 10000
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
config_yaml: /gpfs/wangzixuan/Octopus/workspace/starVLA/examples/Franka/train_files/starvla_cotrain_franka_dual.yaml
output_dir: /gpfs/wangzixuan/Octopus/workspace/starVLA/results/Checkpoints/0615_franka_v3_dual_finetune_qwen3OFT