2024-05-17 01:02:00 +08:00
|
|
|
### model
|
2024-05-06 21:47:00 +08:00
|
|
|
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
|
|
|
|
|
|
2024-05-17 01:02:00 +08:00
|
|
|
### method
|
2024-05-06 21:47:00 +08:00
|
|
|
stage: pt
|
|
|
|
|
do_train: true
|
|
|
|
|
finetuning_type: lora
|
2024-06-06 03:53:28 +08:00
|
|
|
lora_target: all
|
2024-05-06 21:47:00 +08:00
|
|
|
|
2024-05-17 01:02:00 +08:00
|
|
|
### dataset
|
2024-05-06 21:47:00 +08:00
|
|
|
dataset: c4_demo
|
|
|
|
|
cutoff_len: 1024
|
|
|
|
|
max_samples: 1000
|
|
|
|
|
overwrite_cache: true
|
|
|
|
|
preprocessing_num_workers: 16
|
|
|
|
|
|
2024-05-17 01:02:00 +08:00
|
|
|
### output
|
2024-07-13 13:16:22 +08:00
|
|
|
output_dir: saves/llama3-8b/lora/pretrain
|
2024-05-06 21:47:00 +08:00
|
|
|
logging_steps: 10
|
|
|
|
|
save_steps: 500
|
|
|
|
|
plot_loss: true
|
|
|
|
|
overwrite_output_dir: true
|
|
|
|
|
|
2024-05-17 01:02:00 +08:00
|
|
|
### train
|
2024-05-06 21:47:00 +08:00
|
|
|
per_device_train_batch_size: 1
|
|
|
|
|
gradient_accumulation_steps: 8
|
2024-06-03 19:12:29 +08:00
|
|
|
learning_rate: 1.0e-4
|
2024-05-06 21:47:00 +08:00
|
|
|
num_train_epochs: 3.0
|
|
|
|
|
lr_scheduler_type: cosine
|
2024-06-03 19:12:29 +08:00
|
|
|
warmup_ratio: 0.1
|
2024-06-28 01:17:07 +08:00
|
|
|
bf16: true
|
2024-06-13 03:15:06 +08:00
|
|
|
ddp_timeout: 180000000
|
2024-05-06 21:47:00 +08:00
|
|
|
|
2024-05-17 01:02:00 +08:00
|
|
|
### eval
|
2024-05-13 20:39:36 +08:00
|
|
|
val_size: 0.1
|
2024-05-06 21:47:00 +08:00
|
|
|
per_device_eval_batch_size: 1
|
2024-06-06 01:49:20 +08:00
|
|
|
eval_strategy: steps
|
2024-05-06 21:47:00 +08:00
|
|
|
eval_steps: 500
|
