---
license: cc-by-4.0
task_categories:
  - text-classification
language:
  - zh
tags:
  - llm-safety
  - over-refusal
  - false-refusal
  - cybersecurity
  - dual-use
  - benchmark
  - defensive-security
pretty_name: Defensive Refusal Bench (zh) — cyber over-refusal
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: test.jsonl
---

# defensive-refusal-bench-zh · 中文网安「防御问题误拒」评测基准

测安全对齐模型是否把**正当防御/安全教育问题**也误拒（over-refusal），用**真实危害请求**作对照防刷分。与 [agent-safety-bench-zh](https://github.com/uninhibited-scholar/agent-safety-bench-zh) 互补："该拦的拦没拦住" vs "不该拒的有没误拒"。

- v0.2 共 **134 条**：112 `should_answer`（从 [cybersec-qa-dataset-zh](https://huggingface.co/datasets/uninhibited-scholar/cybersec-qa-dataset-zh) 语料挖的边界问题）+ 22 `should_refuse`（实施性危害对照）。
- 字段：`id, label, category, difficulty, question, gold{action}, why, surface_cues, tags`。

## 加载
```python
from datasets import load_dataset
ds = load_dataset("uninhibited-scholar/defensive-refusal-bench-zh", split="test")
```

## 基线（朴素关键词护栏）
```json
{
  "missing": 0,
  "false_refusal_rate": 1.0,
  "helpfulness_on_legit": 0.0,
  "harmful_refusal_rate": 0.682,
  "safety_utility_gap": 0.682
}
```
**看点**：误拒 100% 的正当防御问题，却仍漏过约 32% 的真实危害——关键词过滤分不清"讲原理"和"搞破坏"。

代码 / 评分器 / CI：<https://github.com/uninhibited-scholar/defensive-refusal-bench-zh>　诚实定位：v0.2、单人构建、规则基线。CC BY 4.0。