zephyr-7B-alpha-dare-0.85: benchmark results
A community DARE (Drop And REscale) pruning experiment on Zephyr 7B Alpha, randomly zeroing 85% of delta weights with rescaling to test parameter redundancy. Provider: uukuguy. Released 2023-11-23. Access: Open.
Aggregate Rating 1363 ± 19, rank #1322 of 1635 rated models, from 20 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-IFEval | 35.24 | Strict accuracy, prompt/instruction mean (%) | 72.5 |
| Open LLM Leaderboard v1 - MMLU | 64.3 | Accuracy (%) (5-shot) | 67.9 |
| Open LLM Leaderboard v1 - WinoGrande | 78.45 | Accuracy (%) (5-shot) | 63.3 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 26.61 | Exact match, flexible extract (%) | 60.6 |
| Open LLM Leaderboard v1 - HellaSwag | 83.67 | Normalized accuracy (%) (10-shot) | 60.6 |
| Open LLM Leaderboard v1 - GSM8K | 42.08 | Accuracy (%) (5-shot) | 56.8 |
| Open Portuguese LLM - OAB Exams | 43.6 | Accuracy (%) | 56 |
| Open Portuguese LLM - FaQuAD NLI | 67.32 | Macro F1 (%) | 53.7 |
| Open LLM Leaderboard v1 - ARC Challenge | 61.18 | Normalized accuracy (%) (25-shot) | 51.6 |
| Open Portuguese LLM - BLUEX | 52.02 | Accuracy (%) | 51.3 |
| Open Portuguese LLM - ENEM | 61.44 | Accuracy (%) | 49.4 |
| Open Korean LLM - KorNAT-Helpful | 46.45 | Normalized accuracy (%) | 46.5 |
Interactive version: theaggregate.ai/model?slug=zephyr-7b-alpha-dare-0-85 · How It Works · Data refreshed daily, snapshot 2026-11-05.