blob: e91176b148ceb11c52224448e30fe07bdc877b27 (
plain)
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
|
### Qwen2.5-1.5B Full SFT Config
model_name_or_path: Qwen/Qwen2.5-1.5B-Instruct
stage: sft
do_train: true
finetuning_type: full
dataset: preference_extractor_train
template: qwen
cutoff_len: 1024
overwrite_cache: true
preprocessing_num_workers: 16
output_dir: saves/qwen2.5-1.5b-full-sft
logging_steps: 10
save_strategy: steps
save_steps: 500
plot_loss: true
overwrite_output_dir: true
per_device_train_batch_size: 8
gradient_accumulation_steps: 16
learning_rate: 2.0e-5
num_train_epochs: 1.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
flash_attn: fa2
val_size: 0.01
per_device_eval_batch_size: 8
eval_strategy: steps
eval_steps: 500
|
