Noromaid-v0.4-Mixtral-Instruct-8x7B-Zloss: benchmark results
An experimental Mixtral Instruct role-play derivative by NeverSleep, credited to Undi and IkariDev. It uses a Z-loss training fork and a weighted LoRA merge below the usual unit weight; the card requests ChatML formatting without special ChatML tokens. Provider: NeverSleep. Access: Open.
Aggregate Rating 1474 ± 33, rank #2750 of 10916 rated models, from 6 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - MMLU | 71.27 | Accuracy (%) (5-shot) | 94.9 |
| Open LLM Leaderboard v1 - HellaSwag | 87.94 | Normalized accuracy (%) (10-shot) | 88.8 |
| Open LLM Leaderboard v1 - ARC Challenge | 70.48 | Normalized accuracy (%) (25-shot) | 86.1 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 66.45 | MC2 (%) (0-shot) | 84.9 |
| Open LLM Leaderboard v1 - WinoGrande | 81.69 | Accuracy (%) (5-shot) | 80.5 |
| Open LLM Leaderboard v1 - GSM8K | 61.49 | Accuracy (%) (5-shot) | 74.2 |
Interactive version: theaggregate.ai/model?slug=noromaid-v0-4-mixtral-instruct-8x7b-zloss · How It Works · Data refreshed daily, snapshot 2026-11-08.