Nejumi 4 - Toxicity - Violation Categories: leaderboard
Japanese safety prompts on questions an assistant should decline or answer without specifics (category 違反カテゴリ) from the toxicity suite of Nejumi Leaderboard 4 (Weights & Biases Japan). GPT-4.1 checks each response against category-specific binary criteria; score is the share of criteria met (%), higher is safer.
Metric: Criteria met (%). Source: nejumi.ai. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 (Thinking) | 61.9 | #148 (Claude Sonnet 4.5) |
| 2 | Claude Opus 4.1 (Thinking) | 61.11 | #146 (Claude Opus 4.1) |
| 3 | GPT-4o (2024-11-20) | 57.94 | #392 |
| 4 | gemma-4-E4B-it | 57.94 | #577 |
| 5 | Claude Haiku 4.5 (Thinking) | 57.94 | #287 (Claude Haiku 4.5) |
| 6 | Nova 2.0 Lite (Non-reasoning) | 57.14 | #425 (Nova 2.0 Lite) |
| 7 | Llama-3.3-Swallow-70B-Instruct-v0.4 | 57.14 | #540 |
| 8 | Claude Sonnet 4 (Thinking) | 56.35 | #207 (Claude Sonnet 4) |
| 9 | Qwen 3.6 35B A3B (Thinking) | 55.56 | #238 (Qwen 3.6 35B A3B) |
| 10 | Claude 3.7 Sonnet (Non-reasoning) | 55.56 | #259 (Claude 3.7 Sonnet) |
| 11 | Gemma 4 E2B (IT) | 54.76 | #779 |
| 12 | Claude 3.7 Sonnet (Thinking) | 54.76 | #259 (Claude 3.7 Sonnet) |
| 13 | mistral-large-4-oss (High) | 54.76 | #60 (mistral-large-4-oss) |
| 14 | Claude 3.5 Sonnet | 53.17 | #362 |
| 15 | Claude Opus 4.5 (Thinking) | 53.17 | #87 (Claude Opus 4.5) |
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-toxicity-violation-categories · How It Works · Data refreshed daily, snapshot 2026-11-05.