Experiment21-7B: benchmark results
A 32-layer Mistral-family checkpoint released as Experiment21-7B. The exact evaluated configuration uses 32,000 vocabulary entries; the release does not document a reproducible training recipe or establish equivalence with other numbered experiments. Provider: Other. Access: Open.
Aggregate Rating 1460 ± 43, rank #3328 of 10916 rated models, from 6 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 79.79 | MC2 (%) (0-shot) | 99.7 |
| Open LLM Leaderboard v1 - WinoGrande | 85.48 | Accuracy (%) (5-shot) | 98.4 |
| Open LLM Leaderboard v1 - HellaSwag | 89.03 | Normalized accuracy (%) (10-shot) | 96.1 |
| Open LLM Leaderboard v1 - ARC Challenge | 71.42 | Normalized accuracy (%) (25-shot) | 89.2 |
| Open LLM Leaderboard v1 - GSM8K | 65.96 | Accuracy (%) (5-shot) | 81 |
| Open LLM Leaderboard v1 - MMLU | 63.92 | Accuracy (%) (5-shot) | 64.1 |
Interactive version: theaggregate.ai/model?slug=experiment21-7b · How It Works · Data refreshed daily, snapshot 2026-11-08.