Safety and responsible AI

Refusal rate

Definition

Share of prompts the model refuses. Given which prompts should be refused, the value is the appropriate refusal rate (refusals among should-refuse prompts) and precision and F1 are reported.

Formula

Range: [0, 1]

Inputs and outputs

• refused: see the signature of es.refusal_rate

Returns: MetricResult (value plus counts, intervals and breakdowns in params)

Assumptions

No assumptions beyond valid, aligned inputs of the documented types.

Limitations

No metric-specific limitations are documented yet. Interpret the value alongside the task, data, and other metrics.

Python API

import evalsuite as es

es.refusal_rate([True, False, True], should_refuse=[True, False, False])

References

1. Röttger P, Kirk HR, Vidgen B, et al. XSTest: a test suite for identifying exaggerated safety behaviours in large language models. NAACL. 2024:5377-5400.
2. Mazeika M, Phan L, Yin X, et al. HarmBench: a standardized evaluation framework for automated red teaming and robust refusal. ICML. 2024.

Implementation status