zephyr-7B-beta: benchmark results
HuggingFace H4's Zephyr-7B-beta, the second Zephyr model, a Mistral 7B fine-tune DPO-aligned on synthetic datasets with built-in safety alignment deliberately removed to boost helpfulness. Provider: HuggingFace. Released 2023-10-27. Access: Open.
Aggregate Rating 1372 ± 10, rank #1330 of 1656 rated models, from 169 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HumanLikeness - Word-2 | 25.64 | Humanlike Score (%) | 100 |
| Open Medical LLM - PubMedQA | 76.6 | Accuracy (%) | 85.4 |
| Hallucinations Leaderboard | 49.62 | Average Task Score (score) | 84.2 |
| HumanLikeness - Meaning-2 | 72.93 | Humanlike Score (%) | 84.2 |
| CyberBench | 57.7 | Average Score (percentage) | 75 |
| CyberBench (NLP) | 57.7 | Published ten-dataset average (%) | 75 |
| AlpacaEval 1.0 | 90.6 | Win Rate (%) | 74.3 |
| InfiBench | 46.31 | Score (%) | 73.3 |
| Open Korean LLM - KorNAT-Helpful | 48.8 | Normalized accuracy (%) | 72.8 |
| CyberMetric | 76.4 | 500-question accuracy (%) | 70.8 |
| LLM Trustworthy - Privacy | 84.18 | Trust Score (%) | 70.8 |
| LatamBoard - FLORES Bidirectional | 40.87 | Score (%) | 69.7 |
Interactive version: theaggregate.ai/model?slug=zephyr-7b-beta · How It Works · Data refreshed daily, snapshot 2026-10-20.