是否有qwen3 8B 在terminal rl 上的训练曲线？

[tAnGjIa520]

[xuyangc03]

[sihan-zzz]

Hi @xuyangc03 thanks for sharing the figure! Wondering what's the truncation ratio curve along the training? we tried to reproduce this work using same qwen3-8b and other configs as in the terminal-rl_qwen3-8b.sh, but we found that the max_turn=10 seems quite tight for qwen3-8b that most of its trials would be truncated without any positive reward. And the truncation ratio kept increasing as training collapse.

Appreciate it if you can share some insights here. Thanks!

[xuyangc03]

Hi @sihan-zzz , thanks for the comment! We observe a similar trend in the truncation ratio of samples. An increasing truncation ratio from early in training is expected, since as RL progresses the model tends to learn to spend more interactions working on a task, so more trajectories eventually hit the iteration budget. There is still a non-zero fraction of non-truncated trajectories, which provides the training signal. The optimal iteration budget can depend on task difficulty. In this setting, max_turn=10 is likely too tight, and it is reasonable to increase it.