[Submission] Thanh 512

[Luvata]

Student Name

Thanh

Model Length

512

Accuracy

60.90%

Improvement Description

Default config, with vLLM logprobs correction

Detailed Write-up

Based on this one
https://fengyao.notion.site/off-policy-rl

And based on the implementation in TRL to get the log probs from vllm and important sampling

https://github.com/huggingface/trl/blob/e086f073cf6dee30acc2d3fe357db21e1901c2be/trl/trainer/grpo_trainer.py#L1258

240 steps 1xh100

GPU Hours

1.5 hour H100

Submission Agreement

• I confirm that these results are from my own work

[github-actions]

✅ Your submission has been processed and added to the leaderboard! View it at: https://newturing.github.io/leaderboard/

[github-actions]

✅ Your submission has been processed and added to the leaderboard! View it at: https://newturing.github.io/leaderboard/