custom_eval:
  comparisons_per_task: 5
  confusion_matrix:
  - aliangdw_metaworld_metaworld_eval
  eval_types:
  - reward_alignment
  - policy_ranking
  num_examples_per_quality_pr: 5
  policy_ranking:
  - aliangdw_metaworld_metaworld_eval
  - aliangdw_utd_so101_policy_ranking_utd_so101_policy_ranking
  - aliangdw_usc_franka_policy_ranking_usc_franka_policy_ranking
  - aliangdw_usc_xarm_policy_ranking_usc_xarm_policy_ranking
  - jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist
  quality_preference:
  - mw
  reward_alignment:
  - jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist
  - aliangdw_metaworld_metaworld_eval
  similarity_score:
  - aliangdw_metaworld_metaworld_eval
data:
  data_source_weights:
    metaworld_train: 1.0
    roboarena: 1.0
  dataloader_num_workers: 8
  dataloader_persistent_workers: true
  dataloader_pin_memory: true
  dataset_preference_ratio: 0.7
  dataset_success_cutoff_file: rfm/data/dataset_success_cutoff.txt
  dataset_type: rfm
  eval_datasets:
  - jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist
  eval_subset_size: null
  fps: 10
  load_embeddings: true
  max_frames: 8
  max_frames_after_preprocessing: 64
  max_success: 1.0
  max_trajectories: -1
  min_frames_per_trajectory: 10
  min_success: 0.8
  n_wrong_tasks: 5
  num_bins: 10
  pairwise_progress: false
  preference_strategy_ratio:
  - 6.0
  - 1.0
  - 1.0
  progress_pred_type: absolute_wrt_total_frames
  progress_strategy_ratio:
  - 1.0
  - 2.0
  - 1.0
  - 1.0
  - 1.0
  resized_height: 196
  resized_width: 196
  rewind_lengths: null
  roboarena_partial_success_threshold: 0.2
  sample_type_ratio:
  - 1.0
  - 0.0
  - 0.0
  samples_per_trajectory: 1
  seed: 42
  shuffle: true
  shuffle_progress_frames: false
  similarity_strategy_ratio:
  - 1.0
  - 1.0
  - 1.0
  task_instruction_same_source_prob: 0.5
  train_datasets:
  - jesbu1_roboarena_eval_debug_nowrist_roboarena_eval_debug_nowrist
  use_multi_image: true
  use_uniform_sampling: false
debug: false
logging:
  log_level: INFO
  log_to:
  - wandb
  save_best:
    greater_is_better:
    - true
    - true
    hub_private: false
    hub_save_every: 250
    hub_token: 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
    keep_top_k: 5
    metric_names:
    - eval_rew_align/pearson_mw_eval
    - eval_p_rank/spearman_mw_eval
    save_every: 250
    upload_to_hub: false
  save_model: true
  save_processor: true
  wandb_entity: clvr
  wandb_mode: null
  wandb_notes: training RFM
  wandb_project: rfm
loss:
  predict_last_frame_progress: false
  success_positive_weight: 6.0
mode: train
model:
  average_temporal_patches: false
  base_model_id: rewind_scale_transformer
  causal_mask: false
  model_type: default
  pairwise_progress: false
  peft_vision_encoder: false
  quantization: false
  rewind: null
  rewind_scale_model: true
  torch_dtype: bfloat16
  train_language_model: false
  train_preference_head: true
  train_progress_head: true
  train_similarity_head: false
  train_success_head: false
  train_vision_encoder: false
  trust_remote_code: true
  use_multi_image: true
  use_peft: false
  use_progress_token: false
  use_unsloth: false
peft:
  bias: none
  lora_alpha: 64
  lora_dropout: 0.05
  r: 32
  target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj
trainer_cls: rewind_scale_transformer
training:
  beta: 0.1
  bf16: true
  custom_eval_steps: 250
  dataloader_num_workers: 8
  dataloader_persistent_workers: true
  dataloader_pin_memory: true
  ddp_bucket_cap_mb: 25
  ddp_find_unused_parameters: false
  do_eval: true
  eval_steps: 250
  evaluation_strategy: steps
  exp_name: rewind_scale_Progress_Pref_test_save
  fp16: false
  gradient_accumulation_steps: 1
  gradient_checkpointing: false
  learning_rate: 2.0e-05
  logging_steps: 1
  lr_scheduler_type: cosine
  max_grad_norm: 10.0
  max_seq_length: 1024
  max_steps: 100000
  num_gpus: 2
  num_train_epochs: -1
  output_dir: ./rewind_logs
  overwrite_output_dir: true
  per_device_eval_batch_size: 16
  per_device_train_batch_size: 1024
  predict_pref_progress: true
  predict_sim_progress: true
  prediction_loss_only: true
  remove_unused_columns: false
  resume_from_checkpoint: null
  run_default_eval: false
  save_steps: 250
  save_strategy: 'no'
  vision_encoder_lr: 1.0e-05
  vision_encoder_num_layers: 3
  warmup_ratio: 0.1
  warmup_steps: 0
  weight_decay: 0.1