The performance curves obtained during training do not match the results from the actual evaluation.

[arrebolLee]

I reproduced the training process of Prime Verifier using your code and obtained curves similar to those in the paper. However, when I used the trained model to generate answers on the dataset and evaluated them, the measured accuracy differed noticeably from the accuracies shown in those curves—especially on the Olympiad_bench dataset. What could be causing this discrepancy?

[zhangchen-xu]

Hi,

Thank you for your message. This is possibly due to different inference engines. Are you using the same vllm version and greedy decoding?

[arrebolLee]

my vllm version is 0.8.5.post1 and I set temperature = 0 which equals greedy decoding.