| datasets: | |
| vla_data: | |
| action_mode: abs | |
| data_mix: robodojo_v21_all_h50_q99 | |
| data_root_dir: ./playground/Datasets | |
| dataset_py: lerobot_datasets | |
| obs_image_size: | |
| - 224 | |
| - 224 | |
| per_device_batch_size: 16 | |
| video_backend: torchvision_av | |
| framework: | |
| action_model: | |
| action_dim: 14 | |
| action_hidden_dim: 2560 | |
| action_horizon: 50 | |
| action_model_type: MLP | |
| name: QwenOFT | |
| qwenvl: | |
| attn_implementation: flash_attention_2 | |
| base_vlm: Qwen/Qwen3-VL-4B-Instruct | |
| output_dir: ./playground/Checkpoints/robodojo_v21_all_h50_q99_qwenoft | |
| run_id: robodojo_v21_all_h50_q99_qwenoft | |
| run_root_dir: ./playground/Checkpoints | |
| seed: 42 | |
| trainer: | |
| eval_interval: 100 | |
| freeze_modules: '' | |
| gradient_clipping: 1.0 | |
| is_resume: true | |
| learning_rate: | |
| action_model: 0.0001 | |
| base: 1.0e-05 | |
| qwen_vl_interface: 1.0e-05 | |
| logging_frequency: 10 | |
| lr_scheduler_type: cosine_with_min_lr | |
| max_train_steps: 150000 | |
| 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: your_wandb_entity | |
| wandb_project: llavavla | |