haoranli-ml/

Qwen3-4B-Instruct_rl_64_32_nokl_noentropy

nnodes=2
n_gpus_per_node=4
train_batch_size=64
ppo_mini_batch_size=32
rollout_n=8
learning_rate=1e-6
kl_loss_coef=0
entropy_coeff=0
temperature=0.8
clip_ratio_low=0.2
clip_ratio_high=0.26
epoch=10
# below is settings for Dr.GRPO
loss_agg_mode="seq-mean-token-sum-norm"
use_kl_loss=False
norm_adv_by_std_in_grpo=False