Reproducing MATH numbers from OLMo-2 release blog

Thank you for this awesome repo! Seems like a really useful extension of the lm-evaluation-harness. I had a couple of questions regarding the scores from the [OLMo-2 blog](https://allenai.org/blog/olmo2). In particular, under the section "Making OLMo 2 Instruct", the performance of OLMo-2-7B-1124-Instruct on MATH is cited as 31.3.

- Do I understand correctly that this resolves to the minerva_math benchmark?
- Which command can I use to reproduce this? Of all the possible combinations, I got the highest score using the tulu setup, ie. `olmes --model allenai/OLMo-2-1124-7B-Instruct --task "minerva_math::tulu" --output-dir olmes_results/tulu`, shown below:

Summary of primary scores:
minerva_math::tulu: 0.288752
minerva_math_algebra::tulu: 0.506318
minerva_math_counting_and_probability::tulu: 0.291139
minerva_math_geometry::tulu: 0.212944
minerva_math_intermediate_algebra::tulu: 0.125138
minerva_math_number_theory::tulu: 0.227778
minerva_math_prealgebra::tulu: 0.531573
minerva_math_precalculus::tulu: 0.126374

The average of 28.8 is lower than 31.3, so I was wondering if I was doing something wrong here? Furthermore, on a machine with 2x RTXA6000, I get a total processing_time of 108829.94449687004, ie. a bit over 30 hours. Does that seem normal?