# Unsafebench Eval

> Evaluates the effectiveness of image safety classifiers in detecting various unsafe content categories across real-world and AI-generated images. It also probes classifier robustness to distribution shifts caused by artistic representations and grid layouts in AI-generated content. Use when the user wants to benchmark on UnsafeBench, or asks about evaluating this task. Reports F1-Score.

- Skill: `qhjqhj00/unsafebench-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/unsafebench-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/unsafebench-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/unsafebench-eval

---


# unsafebench-eval

> UnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images — Qu et al. (2024) (arXiv:2405.03486, 2024)

## What this evaluates

Evaluates the effectiveness of image safety classifiers in detecting various unsafe content categories across real-world and AI-generated images. It also probes classifier robustness to distribution shifts caused by artistic representations and grid layouts in AI-generated content.

## Datasets

- **UnsafeBench** — total ?; splits: LAION-5B (Real-World) (-1), Lexica (AI-Generated) (-1)

## Metrics

- `F1-Score` **(primary)** — range: [0, 1]
  - Harmonic mean of precision and recall, calculated per unsafe category and overall. Precision = TP / (TP + FP), Recall = TP / (TP + FN).

## Input / output format

**Input**: Single images sourced from LAION-5B (real-world) or Lexica (AI-generated), optionally accompanied by text prompts for VLM-based classifiers.

**Output**: Binary or multi-class safety labels (e.g., safe, hate, harassment, violence, sexual, shocking, illegal activity, deception, political, health, spam) per image.

## Scoring recipe

```python
def compute_f1(predictions, gold_labels):
    tp = sum(1 for p, g in zip(predictions, gold_labels) if p == 1 and g == 1)
    fp = sum(1 for p, g in zip(predictions, gold_labels) if p == 1 and g == 0)
    fn = sum(1 for p, g in zip(predictions, gold_labels) if p == 0 and g == 1)
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    return 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
```

## Common pitfalls

- Relying on a single prompt for VLM-based classifiers without majority voting, which can drop F1-Score by ~0.10.
- Ignoring distribution shifts between real-world and AI-generated images, leading to overestimation of classifier robustness.
- Overlooking the impact of artistic representations and grid layouts, which cause significant false negatives for models trained only on real-world data.

## Evidence (verbatim from paper)

> We calculate the overall F1-Score of Q16 combined with NudeNet across the unsafe categories they can cover, and the score is 0.665.

## Citation

```bibtex
@misc{qu2024unsafebench,
  title={UnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images},
  author={Qu et al. (2024)},
  year={2024},
  note={arXiv:2405.03486}
}
```

- arXiv: 2405.03486