Datasets:

Modalities:
Image
ArXiv:
License:
SZhanZ commited on
Commit
90d57b2
·
verified ·
1 Parent(s): 9b1b129

Upload qwen2_5_vl_full_sft.yaml

Browse files
sft_related/qwen2_5_vl_full_sft.yaml CHANGED
@@ -1,6 +1,6 @@
1
  ### model
2
  model_name_or_path: Qwen/Qwen2.5-VL-3B-Instruct
3
- image_max_pixels: 262144
4
  video_max_pixels: 16384
5
  trust_remote_code: true
6
 
@@ -8,16 +8,16 @@ trust_remote_code: true
8
  stage: sft
9
  do_train: true
10
  finetuning_type: full
11
- freeze_vision_tower: true # choices: [true, false]
12
- freeze_multi_modal_projector: true # choices: [true, false]
13
  train_mm_proj_only: false # choices: [true, false]
14
  deepspeed: examples/deepspeed/ds_z2_config.json # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]
15
 
16
  ### dataset
17
  dataset: mllm_rec_json
18
  template: qwen2_vl
19
- cutoff_len: 2048
20
- max_samples: 1000000
21
  overwrite_cache: true
22
  preprocessing_num_workers: 16
23
 
@@ -31,8 +31,8 @@ overwrite_output_dir: true
31
  ### train
32
  per_device_train_batch_size: 1
33
  gradient_accumulation_steps: 2
34
- learning_rate: 2.0e-5
35
- num_train_epochs: 3.0
36
  lr_scheduler_type: cosine
37
  warmup_ratio: 0.1
38
  bf16: true
 
1
  ### model
2
  model_name_or_path: Qwen/Qwen2.5-VL-3B-Instruct
3
+ image_max_pixels: 12845056
4
  video_max_pixels: 16384
5
  trust_remote_code: true
6
 
 
8
  stage: sft
9
  do_train: true
10
  finetuning_type: full
11
+ freeze_vision_tower: false # choices: [true, false]
12
+ freeze_multi_modal_projector: false # choices: [true, false]
13
  train_mm_proj_only: false # choices: [true, false]
14
  deepspeed: examples/deepspeed/ds_z2_config.json # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]
15
 
16
  ### dataset
17
  dataset: mllm_rec_json
18
  template: qwen2_vl
19
+ cutoff_len: 131072
20
+ max_samples: 10000000
21
  overwrite_cache: true
22
  preprocessing_num_workers: 16
23
 
 
31
  ### train
32
  per_device_train_batch_size: 1
33
  gradient_accumulation_steps: 2
34
+ learning_rate: 1.0e-6
35
+ num_train_epochs: 2.0
36
  lr_scheduler_type: cosine
37
  warmup_ratio: 0.1
38
  bf16: true