MATH (Maj1@4)
Math · 2021-03-05
MATH accuracy with majority voting over 4 sampled solutions (maj@4), usually with 4-shot Minerva prompting, on the 5,000-problem test set. Kept separate from single-sample MATH.
Top models (higher is better)
| Model | Score |
|---|---|
| Mistral Small 3 24B Base 2501 | 46.0 |
| Mixtral 8x22B Instruct | 44.6 |
| Mixtral 8x22B | 41.8 |
| Mixtral 8x7B | 28.4 |
| Llama 2 70B | 13.8 |
| LLaMA 33B | 8.4 |
| Llama 2 13B | 6.0 |
| Llama 2 7B | 3.9 |