LLaMA-Factory/examples/train_lora/llama3_lora_ppo.yaml

40 lines
710 B
YAML
Raw Normal View History

2024-05-16 17:02:00 +00:00
### model
2024-05-06 13:47:00 +00:00
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
reward_model: saves/llama3-8b/lora/reward
2024-05-16 17:02:00 +00:00
### method
2024-05-06 13:47:00 +00:00
stage: ppo
do_train: true
finetuning_type: lora
2024-06-05 19:53:28 +00:00
lora_target: all
2024-05-06 13:47:00 +00:00
2024-05-16 17:02:00 +00:00
### dataset
2024-05-17 19:44:56 +00:00
dataset: identity,alpaca_en_demo
2024-05-06 13:47:00 +00:00
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
2024-05-16 17:02:00 +00:00
### output
2024-05-06 13:47:00 +00:00
output_dir: saves/llama3-8b/lora/ppo
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
2024-05-16 17:02:00 +00:00
### train
2024-05-06 13:47:00 +00:00
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
2024-06-03 11:12:29 +00:00
learning_rate: 1.0e-5
2024-05-06 13:47:00 +00:00
num_train_epochs: 3.0
lr_scheduler_type: cosine
2024-06-03 11:12:29 +00:00
warmup_ratio: 0.1
2024-06-27 17:17:07 +00:00
bf16: true
2024-06-12 19:15:06 +00:00
ddp_timeout: 180000000
2024-05-06 13:47:00 +00:00
2024-05-16 17:02:00 +00:00
### generate
2024-05-06 13:47:00 +00:00
max_new_tokens: 512
top_k: 0
top_p: 0.9