-
Notifications
You must be signed in to change notification settings - Fork 6
Expand file tree
/
Copy pathrun.sh
More file actions
48 lines (48 loc) · 1.22 KB
/
run.sh
File metadata and controls
48 lines (48 loc) · 1.22 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
python verifree.py \
--critic_type rloo \
--gpus 8 \
--enable_prefix_caching \
--collocate \
--vllm_sleep \
--vllm_gpu_ratio 0.45 \
--gradient-checkpointing \
--flash-attn \
--bf16 \
--rnd-seed \
--learning_rate 0.000001 \
--lr_scheduler constant \
--kl_penalty_coef 0 \
--num_ppo_epochs 1 \
--beta 0 \
--non_stop_fixed_reward 0 \
--oracle_type reward \
--oracle math \
--pretrain Qwen/Qwen3-8B-Base \
--prompt_template ours \
--zero-stage 2 \
--ref_offload \
--prompt_data ./datasets/webdata \
--train_split train \
--input_key problem \
--output_key answer \
--max-train 9999999 \
--num_prompt_epoch 999 \
--prompt_max_length 1024 \
--sync_params_every 1 \
--num_samples 8 \
--max_step_adjustment 8 \
--temperature 1 \
--top_p 1 \
--generate_max_length 3000 \
--save_steps 128 \
--max_save_num 999 \
--max_save_mem 99999999 \
--train_batch_size 128 \
--train_batch_size_per_device 1 \
--rollout_batch_size 16 \
--rollout_batch_size_per_device 2 \
--pi_buffer_maxlen_per_device 16 \
--eval_steps -1 \
--reward_source p \
--sft_coef_source reward \
--debug