2024-09-04 16:52:15 +08:00
|
|
|
### model
|
|
|
|
|
model_name_or_path: ../../llm/baichuan
|
|
|
|
|
|
|
|
|
|
### method
|
|
|
|
|
stage: sft
|
|
|
|
|
do_train: true
|
|
|
|
|
finetuning_type: lora
|
|
|
|
|
lora_target: all
|
|
|
|
|
|
|
|
|
|
### dataset
|
|
|
|
|
dataset: belle_1m
|
|
|
|
|
template: baichuan
|
|
|
|
|
cutoff_len: 1024
|
|
|
|
|
max_samples: 10000
|
|
|
|
|
overwrite_cache: true
|
|
|
|
|
preprocessing_num_workers: 16
|
|
|
|
|
|
|
|
|
|
### output
|
|
|
|
|
output_dir: ./results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1_single_step500
|
|
|
|
|
logging_steps: 3
|
|
|
|
|
save_steps: 500
|
|
|
|
|
plot_loss: true
|
|
|
|
|
overwrite_output_dir: true
|
|
|
|
|
|
|
|
|
|
### train
|
|
|
|
|
per_device_train_batch_size: 2
|
|
|
|
|
gradient_accumulation_steps: 8
|
2024-10-23 16:06:02 +08:00
|
|
|
learning_rate: 1.0e-4
|
2024-09-04 16:52:15 +08:00
|
|
|
num_train_epochs: 10.0
|
|
|
|
|
lr_scheduler_type: cosine
|
|
|
|
|
warmup_ratio: 0.1
|
2024-09-05 15:49:32 +08:00
|
|
|
fp16: true
|
2024-09-04 16:52:15 +08:00
|
|
|
ddp_timeout: 180000000
|
|
|
|
|
max_steps: 500
|
|
|
|
|
include_num_input_tokens_seen: true
|
|
|
|
|
include_tokens_per_second: true
|
|
|
|
|
|
|
|
|
|
### eval
|
|
|
|
|
val_size: 0.1
|
|
|
|
|
per_device_eval_batch_size: 2
|
|
|
|
|
eval_strategy: steps
|
|
|
|
|
eval_steps: 500
|
