| data: | |
| dataloader_num_workers: 0 | |
| dataloader_pin_memory: false | |
| dataset_preference_ratio: 0.7 | |
| dataset_type: default | |
| eval_datasets: | |
| - abraranwar/libero_rfm | |
| - ykorkmaz/libero_failure_rfm | |
| - HenryZhang/metaworld_rewind_rfm_eval | |
| eval_subset_size: 500 | |
| eval_subsets: | |
| - - libero256_10 | |
| - - libero_10_failure | |
| - - metaworld_rewind_eval | |
| force_reprocess: false | |
| fps: 10 | |
| max_frames: 16 | |
| max_frames_for_preprocessing: 64 | |
| max_trajectories: -1 | |
| model_type: default | |
| n_wrong_tasks: 5 | |
| num_bins: 10 | |
| num_proc: 1 | |
| preference_ratio: 1.0 | |
| preference_strategy_ratio: | |
| - 0.4 | |
| - 0.3 | |
| - 0.3 | |
| - 0.0 | |
| progress_ratio: 0.5 | |
| resized_height: 128 | |
| resized_width: 128 | |
| rewind_lengths: null | |
| samples_per_trajectory: 1 | |
| seed: 42 | |
| shuffle: true | |
| train_datasets: | |
| - abraranwar/libero_rfm | |
| - ykorkmaz/libero_failure_rfm | |
| - HenryZhang/metaworld_rewind_rfm_train | |
| train_subsets: | |
| - - libero256_90 | |
| - - libero_90_failure | |
| - - metaworld_rewind_train | |
| video_frame_sampling: uniform | |
| debug: false | |
| logging: | |
| print_trainable_parameters: true | |
| save_model: true | |
| save_processor: true | |
| use_wandb: true | |
| wandb_entity: clvr | |
| wandb_project: rfm | |
| wandb_run_name: rfm | |
| mode: train | |
| model: | |
| base_model_id: Qwen/Qwen2.5-VL-3B-Instruct | |
| torch_dtype: bfloat16 | |
| train_language_model: false | |
| train_preference_head: true | |
| train_progress_head: true | |
| train_similarity_head: false | |
| train_value_head: true | |
| train_vision_encoder: true | |
| trust_remote_code: true | |
| peft: | |
| bias: none | |
| lora_alpha: 64 | |
| lora_dropout: 0.05 | |
| peft_vision_encoder: true | |
| r: 32 | |
| target_modules: | |
| - q_proj | |
| - k_proj | |
| - v_proj | |
| - o_proj | |
| - gate_proj | |
| - up_proj | |
| - down_proj | |
| use_peft: false | |
| training: | |
| beta: 0.1 | |
| bf16: true | |
| ddp_bucket_cap_mb: 25 | |
| ddp_find_unused_parameters: true | |
| do_eval: true | |
| eval_steps: 100 | |
| evaluation_strategy: steps | |
| fp16: false | |
| gradient_accumulation_steps: 1 | |
| gradient_checkpointing: true | |
| learning_rate: 2.0e-05 | |
| logging_steps: 1 | |
| lr_scheduler_type: cosine | |
| max_grad_norm: 10.0 | |
| max_seq_length: 1024 | |
| max_steps: 5000 | |
| num_gpus: 2 | |
| num_train_epochs: -1 | |
| output_dir: ./logs/rfm_progpref_peft_vision | |
| per_device_eval_batch_size: 8 | |
| per_device_train_batch_size: 8 | |
| prediction_loss_only: true | |
| remove_unused_columns: false | |
| resume_from_checkpoint: null | |
| save_steps: 200 | |
| save_strategy: steps | |
| warmup_ratio: 0.1 | |
| warmup_steps: 0 | |