Nejumi 4 - ALT - Toxicity: leaderboard
Prompts from LINE Yahoo's non-public inappropriate speech evaluation set fall into four categories: fairness, social norms, prohibited behaviour and violation category. An LLM judge checks each reply against three yes or no criteria for its category, such as taking a neutral stance. The board is the mean of the four category pass rates in percent, higher is better. Source: github.com/wandb/llm-leaderboard.
Metric: Score (%). Source: nejumi.ai. Saturation forecast: Estimated already saturated. 139 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 88.21 | #280 |
| 2 | Claude Opus 4.1 | 87.79 | #140 |
| 3 | Qwen 3.6 35B A3B | 87.52 | #226 |
| 4 | Claude Sonnet 4.5 | 87.32 | #142 |
| 5 | Qwen 3.6 Max Preview | 87.13 | #85 |
| 6 | Qwen 3.5 27B | 86.43 | #187 |
| 7 | Inkling Small | 86.39 | #138 |
| 8 | Claude Opus 4.5 | 86.35 | #83 |
| 9 | Inkling | 85.96 | #126 |
| 10 | Claude 3.7 Sonnet | 85.78 | #249 |
| 11 | Solar Pro 4 | 85.74 | #215 |
| 12 | Claude Opus 4.8 | 85.68 | #37 |
| 13 | MiniMax-M3 | 85.62 | #136 |
| 14 | DeepSeek V3.2 | 85.6 | #205 |
| 15 | Claude Sonnet 4 | 85.48 | #199 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-alt-toxicity · How It Works · Data refreshed daily, snapshot 2026-10-20.