Hugging Face's logo

amburger66
/
robometer-4b-lora-robotsmith-task02

Transformers
Safetensors
qwen3_vl
reward_model
rbm
preference_comparisons
Model card Files
xet
 Community
amburger66 commited on
Commit
7fc3290
·
verified ·
1 Parent(s): 933e925

Upload config.yaml with huggingface_hub

Browse files
Files changed (1) hide show
  1. config.yaml +177 -0
config.yaml ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ custom_eval:
2
+ comparisons_per_task: 5
3
+ confusion_matrix:
4
+ - mw
5
+ custom_eval_random_seed: 42
6
+ eval_types:
7
+ - reward_alignment
8
+ - policy_ranking
9
+ max_comparisons: null
10
+ num_examples_per_quality_pr: 5
11
+ num_partial_successes: 5
12
+ pad_frames: true
13
+ policy_ranking:
14
+ - amburger66_robotsmith_rbm_task02_robotsmith
15
+ policy_ranking_max_tasks: 100
16
+ quality_preference:
17
+ - mw
18
+ reward_alignment:
19
+ - amburger66_robotsmith_rbm_task02_robotsmith
20
+ reward_alignment_max_trajectories: 10
21
+ subsample_n_frames: null
22
+ use_frame_steps: true
23
+ data:
24
+ data_source_weights: null
25
+ dataloader_num_workers: 8
26
+ dataloader_persistent_workers: true
27
+ dataloader_pin_memory: true
28
+ dataset_preference_ratio: 0.7
29
+ dataset_success_cutoff_file: robometer/data/dataset_success_cutoff.txt
30
+ dataset_type: rbm
31
+ eval_datasets:
32
+ - amburger66_robotsmith_rbm_task02_robotsmith
33
+ eval_subset_size: null
34
+ load_embeddings: false
35
+ max_frames: 16
36
+ max_frames_after_preprocessing: 64
37
+ max_success: 1.0
38
+ max_trajectories: -1
39
+ min_frames_per_trajectory: 5
40
+ min_success: 0.5
41
+ partial_success_threshold: 0.2
42
+ predict_last_frame_partial_progress: false
43
+ preference_strategy_ratio:
44
+ - 1.0
45
+ - 1.0
46
+ - 1.0
47
+ - 1.0
48
+ progress_discrete_bins: 10
49
+ progress_loss_type: discrete
50
+ progress_pred_type: absolute_first_frame
51
+ progress_strategy_ratio:
52
+ - 1.0
53
+ - 1.0
54
+ - 1.0
55
+ - 1.0
56
+ resized_height: null
57
+ resized_width: null
58
+ sample_type_ratio:
59
+ - 1.0
60
+ - 0.0
61
+ - 0.0
62
+ seed: 42
63
+ shuffle: true
64
+ shuffle_progress_frames: false
65
+ train_datasets:
66
+ - amburger66_robotsmith_rbm_task02_robotsmith
67
+ traj_same_source_prob: 0.5
68
+ use_multi_image: true
69
+ use_per_frame_progress_token: true
70
+ debug: false
71
+ logging:
72
+ log_level: INFO
73
+ log_to:
74
+ - wandb
75
+ save_best:
76
+ greater_is_better:
77
+ - true
78
+ - true
79
+ hub_private: false
80
+ hub_save_every: 1000
81
+ hub_token: null
82
+ keep_top_k: 5
83
+ metric_names:
84
+ - eval_rew_align/pearson_amburger66_robotsmith_rbm_task02_robotsmith
85
+ - eval_p_rank/kendall_last_amburger66_robotsmith_rbm_task02_robotsmith
86
+ save_every: 1000
87
+ upload_to_hub: false
88
+ save_model: true
89
+ save_processor: true
90
+ wandb_entity: r-pad
91
+ wandb_mode: null
92
+ wandb_notes: fine-tuning Robometer on RobotSmith
93
+ wandb_project: rbm-finetune-robotsmith
94
+ loss:
95
+ predict_last_frame_progress: false
96
+ progress_discrete_bins: 10
97
+ progress_loss_type: discrete
98
+ success_positive_weight: 1.0
99
+ mode: train
100
+ model:
101
+ average_temporal_patches: true
102
+ base_model_id: Qwen/Qwen3-VL-4B-Instruct
103
+ frame_pooling: mean
104
+ frame_pooling_attn_temperature: 1.0
105
+ model_type: default
106
+ peft_vision_encoder: false
107
+ progress_discrete_bins: 10
108
+ progress_loss_type: discrete
109
+ quantization: false
110
+ rewind: null
111
+ torch_dtype: bfloat16
112
+ train_language_model: true
113
+ train_preference_head: true
114
+ train_progress_head: true
115
+ train_success_head: true
116
+ train_vision_encoder: false
117
+ trust_remote_code: true
118
+ use_multi_image: true
119
+ use_peft: true
120
+ use_per_frame_progress_token: true
121
+ use_unsloth: true
122
+ peft:
123
+ bias: none
124
+ lora_alpha: 64
125
+ lora_dropout: 0.05
126
+ peft_vision_encoder: false
127
+ r: 32
128
+ target_modules:
129
+ - q_proj
130
+ - k_proj
131
+ - v_proj
132
+ - o_proj
133
+ - gate_proj
134
+ - up_proj
135
+ - down_proj
136
+ trainer_cls: rbm_heads
137
+ training:
138
+ beta: 0.1
139
+ bf16: true
140
+ custom_eval_steps: 50
141
+ dataloader_num_workers: 8
142
+ dataloader_persistent_workers: true
143
+ dataloader_pin_memory: true
144
+ ddp_bucket_cap_mb: 25
145
+ ddp_find_unused_parameters: false
146
+ do_eval: true
147
+ eval_steps: 50
148
+ evaluation_strategy: steps
149
+ exp_name: lora_task02
150
+ fp16: false
151
+ gradient_accumulation_steps: 1
152
+ gradient_checkpointing: true
153
+ learning_rate: 2.0e-05
154
+ load_from_checkpoint: robometer/Robometer-4B
155
+ logging_steps: 1
156
+ lr_scheduler_type: cosine
157
+ max_grad_norm: 10.0
158
+ max_seq_length: 1024
159
+ max_steps: 1000
160
+ num_gpus: 2
161
+ num_train_epochs: -1
162
+ output_dir: /data/robometer/logs/task02
163
+ overwrite_output_dir: true
164
+ per_device_eval_batch_size: 16
165
+ per_device_train_batch_size: 8
166
+ predict_pref_progress: true
167
+ prediction_loss_only: true
168
+ remove_unused_columns: false
169
+ resume_from_checkpoint: null
170
+ run_default_eval: false
171
+ save_steps: 200
172
+ save_strategy: 'no'
173
+ vision_encoder_lr: 1.0e-05
174
+ vision_encoder_num_layers: 3
175
+ warmup_ratio: 0.1
176
+ warmup_steps: 0
177
+ weight_decay: 0.01