VARM-Bench - Fine-Grained Category: leaderboard

Metric: Macro-F1 (%) of the fine-grained abuse category (six harmful categories and non-attack); deduplicated 1,600-item test set of Chinese social-media posts and comments; the model writes one rationale with six anchored decisions (target, target type, target explicitness, author stance, harmfulness label, fine-grained category) that a deterministic parser scores, unparseable outputs counting as wrong; deterministic decoding; zero-shot prompt. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.

Top models

#ModelScore
1DeepSeek V4 Pro86.8
2Qwen 3.7 Max86.4
3GPT-5.585.7

Interactive version: theaggregate.ai/benchmark?slug=varm-bench-fine-grained-category · How It Works · Data refreshed daily, snapshot 2026-09-29.