MERA Original Paper: leaderboard
The baseline evaluation printed in the original MERA paper (Fenogenova et al., ACL 2024): Russian-language tasks run with the authors' lm-eval-based harness at benchmark code v1.1.0, with fixed instruction prompts, a fixed number of shots per task, log-likelihood scoring for classification and multiple-choice tasks and greedy generation otherwise. The total score is the mean of the 17 problem-solving and exam-based task scores in percent, higher is better; these runs are separate from the live MERA leaderboard, whose submissions use later code versions.
Metric: Total Score (%; mean of 17 problem-solving and exam tasks, MERA code v1.1.0). Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral-7B-v0.1 | 40 | #6899 |
| 2 | davinci-002 | 38.3 | #7916 |
| 3 | Llama 2 13B Base | 36.8 | #7995 |
| 4 | Yi 6B (Base) | 35.4 | #8358 |
| 5 | Llama 2 7B Base | 32.7 | #8986 |
| 6 | ruGPT-3.5 | 20.8 | #9589 |
| 7 | ruGPT-3-medium | 20.1 | #9690 |
| 8 | mGPT | 19.8 | #10733 |
| 9 | mGPT-13B | 19.6 | #9566 |
| 10 | ruGPT3-large | 19.3 | #10217 |
| 11 | ruT5-base | 19.3 | #9694 |
| 12 | ruGPT-3-small | 19.1 | #9600 |
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 · How It Works · Data refreshed daily, snapshot 2026-11-12.