zephyr-7B-alpha: benchmark results
HuggingFace H4's Zephyr-7B-alpha, a Mistral 7B fine-tune trained on UltraChat then DPO-aligned on GPT-4-ranked UltraFeedback data, with built-in safety alignment deliberately removed to boost helpfulness. Provider: HuggingFace. Released 2023-10-09. Access: Open.
Aggregate Rating 1376 ± 15, rank #1312 of 1656 rated models, from 82 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Hallucinations Leaderboard | 52.08 | Average Task Score (score) | 97.4 |
| HumanLikeness - Discourse-2 | 74.58 | Humanlike Score (%) | 89.5 |
| HumanLikeness - Meaning-2 | 73.13 | Humanlike Score (%) | 89.5 |
| HumanLikeness - Syntax-1 | 84.56 | Humanlike Score (%) | 89.5 |
| HumanLikeness - Word-2 | 22.88 | Humanlike Score (%) | 73.7 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 57.9 | MC2 (%) (0-shot) | 71.2 |
| Open LLM Leaderboard v1 - WinoGrande | 78.61 | Accuracy (%) (5-shot) | 64.3 |
| Open LLM Leaderboard v1 - HellaSwag | 84.04 | Normalized accuracy (%) (10-shot) | 63.3 |
| HumanLikeness - Sound-2 | 61.83 | Humanlike Score (%) | 63.2 |
| Open LLM Leaderboard - IFEval | 51.91 | Score | 61.9 |
| Open Portuguese LLM - FaQuAD NLI | 69.63 | Macro F1 (%) | 59.7 |
| AlpacaEval | 73.47 | Length-Controlled Win Rate (%) | 57.9 |
Interactive version: theaggregate.ai/model?slug=zephyr-7b-alpha · How It Works · Data refreshed daily, snapshot 2026-10-20.