ClinAlign/train.sh at main · AQ-MedAI/ClinAlign · GitHub

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
49
50
51
52
53
54
55
56
57
58
59
60
61
TRAIN_DATA=PATH_FOR_TRAIN_DATA
VAL_DATA=PATH_FOR_VAL_DATA
TRAIN_MODEL=Qwen/Qwen3-4B-Instruct-2507
train_batch_size=64
ppo_mini_batch_size=512
ppo_micro_batch_size_per_gpu=64
actor_rollout_n=8
nnode=8
CKPT_DIR=PATH_FOR_CKPT
reward_func_path=PATH_TO_rubric_checker.py
logdir="/personal/logs/$(date +%Y-%m-%d)"
mkdir -p "$logdir"

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=$TRAIN_DATA \
    data.val_files=$VAL_DATA \
    data.train_batch_size=$train_batch_size \
    data.val_batch_size=1024 \
    data.max_prompt_length=8192 \
    data.max_response_length=12288 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    data.return_raw_chat=True \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.path=$TRAIN_MODEL \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=$ppo_mini_batch_size \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=$ppo_micro_batch_size_per_gpu \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.entropy_coeff=0 \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=$ppo_micro_batch_size_per_gpu \
    actor_rollout_ref.rollout.tensor_model_parallel_size=8 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
    actor_rollout_ref.rollout.max_num_batched_tokens=65536 \
    actor_rollout_ref.rollout.temperature=0.7 \
    actor_rollout_ref.rollout.n=$actor_rollout_n \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=$ppo_micro_batch_size_per_gpu \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger=['console','tensorboard'] \
    trainer.use_rewrite_dpo=False \
    trainer.use_generative_rm=False \
    trainer.project_name=Qwen3-32B \
    trainer.experiment_name=healthrubric \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=$nnode \
    trainer.save_freq=10 \
    trainer.test_freq=-1 \
    trainer.total_epochs=100 \
    trainer.val_before_train=False \
    trainer.default_local_dir=$CKPT_DIR \
    custom_reward_function.path=$reward_func_path \
    custom_reward_function.name=compute_score $@ 2>&1 | tee "$logdir/${EXECUTIONRECORD_ID}.log"