mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
23 lines
726 B
Bash
23 lines
726 B
Bash
accelerate launch \
|
|
--mixed_precision bf16 \
|
|
--num_processes 1 \
|
|
open_instruct/dpo_tune_cache.py \
|
|
--model_name_or_path EleutherAI/pythia-14m \
|
|
--tokenizer_name EleutherAI/pythia-14m \
|
|
--use_slow_tokenizer False \
|
|
--use_flash_attn False \
|
|
--max_seq_length 1024 \
|
|
--per_device_train_batch_size 1 \
|
|
--gradient_accumulation_steps 4 \
|
|
--learning_rate 5e-07 \
|
|
--lr_scheduler_type linear \
|
|
--warmup_ratio 0.1 \
|
|
--weight_decay 0.0 \
|
|
--num_train_epochs 3 \
|
|
--output_dir output/dpo_pythia_14m/ \
|
|
--with_tracking \
|
|
--report_to wandb \
|
|
--logging_steps 1 \
|
|
--dataset_mixer_list allenai/tulu-3-wildchat-reused-on-policy-8b 100 \
|
|
--add_bos \
|
|
--seed 123 |