SAI LabsBook a demo

Safety IndexEvaluationsXSTest safe-prompt refusals

XSTest safe-prompt refusals

Refusal calibration · Over-refusalRetired from the mean (saturated)Run by SAI

How often the model refuses 250 safe prompts that superficially resemble unsafe ones.

Metric
% refused of 250 safe prompts. Lower is better.
Dataset
github.com/paul-rottger/xstest · 250 safe (+200 unsafe) · CC-BY-4.0
Grader
3-class LLM judge (the paper used GPT-4 / string match)
Status
Run by SAI. Measured on 2 models, 2 readings by SAI.
Replicability
public
Notes
Results so far are SAI engine runs on open-weight models; the HELM Safety XSTest score is listed separately.

Results

ModelValueScoreSourceSetting
DeepSeek-R1-Distill-Qwen-7B1.3%n = 228not scoredSAI
Measured by SAI · 2026-10-20
Run by SAI on the open-weight model served locally; graded by azure gpt-4.1-mini.
Full configurationvLLM, thinking on unless noted, no system prompt; research run; any-refusal rate; judge azure gpt-4.1-mini
Qwen3-8B7.7%n = 233not scoredSAI
Measured by SAI · 2026-10-20
Run by SAI on the open-weight model served locally; graded by azure gpt-4.1-mini.
Full configurationvLLM, thinking on unless noted, no system prompt; research run; any-refusal rate (full refusal 2.1%); judge azure gpt-4.1-mini