benchmark result

[ViviqwerAsd]

where is your benchmark result, how does your trained model perform on the publice dataset? it's important for me, thx !!! :)

[dhcode-cpp]

As far as I know, using the model trained by X-R1 based on Qwen2.5-7B-math, MATH500 can reach 53.4%, while the base model is only 49.8%.

Evaluation code：

https://github.com/dhcode-cpp/X-R1/tree/main?tab=readme-ov-file#benchmark-evaluation

[glennccc]

As far as I know, using the model trained by X-R1 based on Qwen2.5-7B-math, MATH500 can reach 53.4%, while the base model is only 49.8%.

Evaluation code：

https://github.com/dhcode-cpp/X-R1/tree/main?tab=readme-ov-file#benchmark-evaluation

Why I use this benchmark.py to eval DeepSeek-R1-Distill-Qwen-32B, the eval acc is only 0.32?

[ViviqwerAsd]

hmm that's a good grade!! I also wanna know about gsm8k!

[dhcode-cpp]

As far as I know, using the model trained by X-R1 based on Qwen2.5-7B-math, MATH500 can reach 53.4%, while the base model is only 49.8%.
Evaluation code：
https://github.com/dhcode-cpp/X-R1/tree/main?tab=readme-ov-file#benchmark-evaluation

Why I use this benchmark.py to eval DeepSeek-R1-Distill-Qwen-32B, the eval acc is only 0.32?

Please confirm the consistency of the conversation template.

X-R1/src/x_r1/benchmark.py

Line 20 in dd6b0e3

from grpo import SYSTEM_PROMPT