SocialBias-Bench: leaderboard
SocialBias-Bench (Rabbi et al., 2026) has 343 human-centred Python coding tasks, such as social benefits, university admissions or employee benefits, whose classes carry task attributes and sensitive demographic attributes; metamorphic tests vary one sensitive attribute at a time, and code whose output changes is biased. Code bias score: the share of executable generated snippets biased on any attribute, in percent over five completions per task at temperature 1.0; lower is better.
Metric: Code bias score (%; share of executable generated snippets whose output changes when any one sensitive attribute changes; 343 human-centred coding tasks, five completions each at temperature 1.0). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | CodeLlama-70B-Instruct-hf | 28.34 | #6548 |
| 2 | Claude 3 Haiku (20240307) | 36.33 | #2320 |
| 3 | GPT-3.5 Turbo (0125) | 60.58 | #3519 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=socialbias-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.