File size: 1,351 Bytes
d180a71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
data_cfgs:
  eval_data_files: null
  eval_datasets: null
  eval_optional_args: []
  eval_size: null
  eval_split: null
  eval_subset: null
  eval_template: null
  train_data_files: q0_80_preference.pt
  train_datasets: /data/align-anything/hantao/data/mm_interp/AA_preference_cosi_new_step10/tokenized
  train_optional_args: []
  train_size: null
  train_split: train
  train_subset: null
  train_template: Chameleon_preference
logger_cfgs:
  cache_dir: null
  log_project: align-anything
  log_run_name: dpo
  log_type: wandb
  output_dir: /data/align-anything/hantao/align-anything/outputs/mm_interp/AA_preference_cosi_new_step10/q0_80_preference
  save_interval: 400.0
model_cfgs:
  model_max_length: 4096
  model_name_or_path: /data/align-anything/hantao/models/chameleon-7b
  trust_remote_code: true
special_tokens: null
train_cfgs:
  adam_betas:
  - 0.9
  - 0.95
  bf16: true
  ds_cfgs: ds_z3_config.json
  epochs: 3.0
  eval_interval: 10
  eval_strategy: epoch
  fp16: false
  freeze_language_model: true
  freeze_mm_proj: true
  freeze_vision_tower: false
  gradient_accumulation_steps: 2.0
  gradient_checkpointing: true
  learning_rate: 1.0e-06
  lr_scheduler_type: cosine
  lr_warmup_ratio: 0.03
  per_device_eval_batch_size: 4.0
  per_device_train_batch_size: 4.0
  regularization: 0.001
  scale_coeff: 0.1
  seed: 42
  weight_decay: 0.01