Evaluation profile

TukaBench

3sub-evals

0.232%Safety weight

0.874%Freedom weight

2components

Weights below are portfolio-specific global index weights.

Model score (lower is better)Predicted score

About this eval

Multilingual and culturally grounded harmful compliance

Included in the behavior ranking.

Measure	Component	Direction by ranking	Safety weight	Freedom weight
afri_jbb_cultural_asrtukabench/tukabench.csv:afri_jbb_cultural_asrMeasures attack success on culturally adapted harmful prompts.	Safety: human_rights_systemic_harm:0.500\|ordinary_harm_misuse_resistance:0.500 · Freedom: user_direction_following:1.000tukabench	Safety: lower · Freedom: higher	0.155%	0.291%
afri_jbb_harm_asrtukabench/tukabench.csv:afri_jbb_harm_asrMeasures attack success on translated harmful prompts in African languages.	Safety: ordinary_harm_misuse_resistance:1.000 · Freedom: user_direction_following:1.000tukabench	Safety: lower · Freedom: higher	0.0381%	0.291%
afrijail_mono_asrtukabench/tukabench.csv:afrijail_mono_asrMeasures attack success on African-authored culturally grounded prompts.	Safety: ordinary_harm_misuse_resistance:1.000 · Freedom: user_direction_following:1.000tukabench	Safety: lower · Freedom: higher	0.0381%	0.291%

Measures attack success on culturally adapted harmful prompts.

Rank	Model	Value	Provenance
1	grok-4.3	10.6	official
2	gpt-5.2	10.9	official
3	deepseek-v3.2	19.2	official
4	llama-4-maverick	22.4	official
5	gpt-4o	27.5	official
6	grok-3	35.6	official

Measures attack success on translated harmful prompts in African languages.

Rank	Model	Value	Provenance
1	gpt-5.2	2.6	official
2	grok-4.3	8.2	official
3	gpt-4o	13.5	official
4	deepseek-v3.2	16.9	official
4	llama-4-maverick	16.9	official
6	grok-3	29.3	official

Measures attack success on African-authored culturally grounded prompts.

Rank	Model	Value	Provenance
1	gpt-5.2	12.2	official
2	grok-4.3	17.7	official
3	deepseek-v3.2	20.1	official
4	llama-4-maverick	22	official
5	gpt-4o	25.7	official
6	grok-3	30.2	official