inspect-case-bench

Inspect AI implementation of CASE-Bench (Context-Aware Safety Evaluation Benchmark; Sun et al., ICML 2025, arXiv:2501.14940), ported from the reference code at BriansIDP/CASEBench.

Content warning. The dataset contains red-teaming queries involving sensitive topics (self-harm, violence, child abuse, illegal activity). It is intended for safety evaluation only.

What it measures

Each of the 900 samples pairs a potentially unsafe query with a formally described context built on Contextual Integrity theory (sender, recipient, transmission principle). The model under test judges whether responding to the query is safe in that context. The verdict is compared against the human majority label derived from safe_rate in the released data (20 annotations per sample).

The binary protocol from the reference compute_score.py is replicated, including its treatment of the child-related crimes category (scored with the unsafe label rather than excluded). A model output that does not contain a parseable safe/unsafe verdict fails closed to unsafe, so malformed verdicts cannot inflate the safe rate.

Metrics

• accuracy — agreement with the human majority label
• safe_accuracy / unsafe_accuracy — accuracy restricted to samples whose human label is safe / unsafe
• macro_f1 — macro-F1 over the two classes (unsafe = positive)

Dataset

data/CASEBench_data.json is fetched at runtime from the upstream repository, pinned by commit SHA 135f221e0e3e4cdc9a7d17288fa6b54c8aa45d0a. A different pinned URL can be passed with -T data_url=....

Usage

uv sync --extra inspect

# small run against a model
uv run inspect eval inspect_case_bench/case_bench.py@case_bench --model hf/Qwen/Qwen2.5-0.5B-Instruct --limit 20

# offline smoke test
uv run inspect eval inspect_case_bench/case_bench.py@case_bench --model mockllm/model --limit 10

Tests

uv run pytest tests/ -q

Citation

@inproceedings{sun2025casebench,
  title={{CASE-Bench}: Context-Aware Safety Benchmark for Large Language Models},
  author={Sun, Guangzhi and Zhan, Xiao and Feng, Shutong and Woodland, Phil and Such, Jose},
  booktitle={International Conference on Machine Learning},
  year={2025}
}