MERA Original Paper - ruMMLU: leaderboard
ruMMLU from the original MERA paper's baseline runs (Fenogenova et al., ACL 2024, code v1.1.0): a Russian counterpart of MMLU with multiple-choice exam questions over 57 subjects and a newly written closed test set, scored 5-shot by log-likelihood. Accuracy in percent, averaged over domains, higher is better.
Metric: Accuracy (%; 5-shot, log-likelihood). Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral-7B-v0.1 | 67.6 | #6899 |
| 2 | davinci-002 | 61.3 | #7916 |
| 3 | Llama 2 13B Base | 56.3 | #7995 |
| 4 | Yi 6B (Base) | 48.7 | #8358 |
| 5 | Llama 2 7B Base | 45.2 | #8986 |
| 6 | ruGPT-3-medium | 27.1 | #9690 |
| 7 | ruGPT-3-small | 26.3 | #9600 |
| 8 | ruGPT-3.5 | 24.6 | #9589 |
| 9 | ruGPT3-large | 24.5 | #10217 |
| 10 | mGPT | 24.1 | #10733 |
| 11 | ruT5-base | 23.7 | #9694 |
| 12 | mGPT-13B | 23.5 | #9566 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mera-original-paper-rummlu · How It Works · Data refreshed daily, snapshot 2026-11-12.