Evaluation profile

RefusalBench

1sub-evals

0.674%Safety weight

0%Freedom weight

2components

Weights below are portfolio-specific global index weights.

Model score (higher is better)Predicted score

About this eval

Refusal calibration on biological research prompts.

Included in the behavior ranking.

Measure	Component	Direction by ranking	Safety weight	Freedom weight
youden_jrefusalbench/refusalbench.csv:youden_jMeasures whether the model distinguishes dual-use biological requests from matched benign requests rather than refusing indiscriminately.	Safety: benign_helpfulness_non_overrefusal:0.500\|ordinary_harm_misuse_resistance:0.500refusalbench/refusalbench.csv	Safety: higher	0.674%	—

Measures whether the model distinguishes dual-use biological requests from matched benign requests rather than refusing indiscriminately.

Rank	Model	Value	Provenance
1	grok-4.20	0.7871	official
2	claude-sonnet-4.6	0.6766	official
3	claude-opus-4.5	0.6709	official
4	claude-opus-4.6	0.6667	official
5	gpt-5.4-mini	0.3787	official
6	gpt-5.5	0.2979	official
7	claude-opus-4.7	0.234	official
8	gemini-3.1-pro-preview	0.1319	official
9	glm-5	0.08511	official
10	minimax-m2.7	0.08085	official
11	qwen3-next-80b-a3b	0.06383	official
12	nemotron-3-super-120b-a12b	0.06383	official
13	kimi-k2.6	0.0383	official
14	gemini-3.1-flash-lite	0.02979	official
15	nova-pro	0.01282	official
16	deepseek-v3.2	0.01277	official
17	llama-3.3-70b-instruct	0.004255	official
18	mistral-large-3	0.004237	official
19	deepseek-r1	-0.004255	official