sugar/training/train.sh

10 lines
369 B
Bash
Raw Normal View History

2024-11-12 10:20:46 +08:00
accelerate launch --config_file config.yaml train.py \
--pretrained_model_name_or_path=2b_sft_model \
--train_dataset_path=../datasets/code.parquet \
--eval_dataset_path=../datasets/eval.parquet \
--max_length=512 \
--num_train_epochs=3 \
--learning_rate=1e-5 \
--output_dir=../checkpoints/py \
--eval_steps=500 \
--save_steps=1000 \