[Evals] Fix MinervaMath dataset YAML for correct evaluation

[QWangCV]

In the MinervaMath dataset svc-huggingface/minerva-math, the solution field contains both the problem-solving process and the final answer. For example:

Start with:
[
s=\alpha f \text {, }
]
where $s$ is the diameter of the image, $f$ the focal length, and $\alpha$ the angular diameter of the planet. For the values given in the problem:
[
s=\frac{45}{3600} \frac{\pi}{180} 7200=\boxed{1.6} \mathrm{~cm}
]

This format leads to an evaluation accuracy of 0, as the answer is embedded within the full solution text. Since svc-huggingface/minerva-math does not provide a separate field for the final answer, I switched to using the math-ai/minervamath dataset, which includes a clearly separated answer field. I also updated the YAML configuration for the MinervaMath dataset accordingly.

After this modification, the accuracy of Qwen/Qwen2.5-Math-7B on MinervaMath improved from 0 to 0.2684, indicating that the bug has been resolved.

[ko120]

I also encountered this problem, and I figured out that it was caused by using the “math” handler instead of the “minervamath” handler in minervamath.yaml (which handles answer extraction). Changing the dataset as you suggested also resolved it.