mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Costa HuangandGitHub
6c6193b7cd
* Add optional r1-style thinking reward * quick change * test * quick change * push latest change * fix * tested it to work * Push changes * push
41 lines
1.4 KiB
Bash
41 lines
1.4 KiB
Bash
python open_instruct/grpo_vllm_thread_ray_gtrl.py \
|
|
--exp_name $exp_name \
|
|
--output_dir /weka/oe-adapt-default/costah/models/$exp_name \
|
|
--dataset_mixer_list ai2-adapt-dev/rlvr_gsm8k_zs 1.0 \
|
|
--dataset_mixer_list_splits train \
|
|
--dataset_mixer_eval_list ai2-adapt-dev/rlvr_gsm8k_zs 1.0 \
|
|
--dataset_mixer_eval_list_splits train \
|
|
--max_token_length 2048 \
|
|
--max_prompt_token_length 2048 \
|
|
--response_length 2048 \
|
|
--number_samples_per_prompt 4 \
|
|
--model_name_or_path meta-llama/Llama-3.1-8B \
|
|
--stop_strings '"</answer>"' \
|
|
--add_r1_style_format_reward \
|
|
--non_stop_penalty False \
|
|
--stop_token eos \
|
|
--penalty_reward_value 0.0 \
|
|
--temperature 0.7 \
|
|
--ground_truths_key ground_truth \
|
|
--chat_template_name r1_simple_chat_postpend_think \
|
|
--sft_messages_key messages \
|
|
--learning_rate 3e-7 \
|
|
--total_episodes 1000000 \
|
|
--deepspeed_stage 3 \
|
|
--per_device_train_batch_size 1 \
|
|
--local_rollout_forward_batch_size 1 \
|
|
--local_mini_batch_size 16 \
|
|
--local_rollout_batch_size 16 \
|
|
--num_epochs 1 \
|
|
--actor_num_gpus_per_node 6 \
|
|
--vllm_tensor_parallel_size 2 \
|
|
--beta 0.01 \
|
|
--apply_verifiable_reward true \
|
|
--seed 3 \
|
|
--num_evals 100 \
|
|
--save_freq 100 \
|
|
--reward_model_multiplier 0.0 \
|
|
--no_try_launch_beaker_eval_jobs \
|
|
--gradient_checkpointing \
|
|
--with_tracking
|