SORRY-Bench/sorry-bench

Benchmark evaluation code for "SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal" (ICLR 2025)

33
/ 100
Emerging

This tool helps AI safety researchers and developers systematically evaluate how well large language models (LLMs) refuse unsafe or inappropriate requests. You input an LLM (either a local model or an API-based one like GPT or Claude) and it outputs detailed compliance rates across 44 safety categories. This allows practitioners to quickly understand an LLM's safety behaviors and identify areas for improvement.

No commits in the last 6 months.

Use this if you need a comprehensive, granular, and efficient way to benchmark the safety refusal capabilities of various large language models, including how they handle diverse linguistic styles and mutations.

Not ideal if you are looking to fine-tune an LLM for safety or to develop new safety refusal techniques, as this tool focuses on evaluation rather than training.

AI-safety LLM-evaluation responsible-AI model-auditing ethics-in-AI
Stale 6m No Package No Dependents
Maintenance 0 / 25
Adoption 9 / 25
Maturity 16 / 25
Community 8 / 25

How are scores calculated?

Stars

77

Forks

5

Language

Jupyter Notebook

License

MIT

Last pushed

Mar 01, 2025

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/SORRY-Bench/sorry-bench"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.