← Evals

Evaluation profile

Microsoft Phi Safety Panels

4sub-evals
0.337%total index weight
2components

Within-component eval weight: Misuse resistance 1.16% · Human rights 1.48%.

Model score (lower is better)Predicted score

About this eval

Harmful-content and jailbreak defect rates

Included in the behavior ranking.

Sub-evals

MeasureComponentDirectionTotal index weightWithin-component weight
harmful_continuationphi-system-card/phi-system-card.csv:harmful_continuationMeasures harmful-content continuation defect rate.ordinary_harm_misuse_resistance:1.000microsoft-phi-safety-panelsLower is better0.0361%Misuse resistance 0.361%
harmful_summarizationphi-system-card/phi-system-card.csv:harmful_summarizationMeasures harmful-content summarization defect rate.ordinary_harm_misuse_resistance:1.000microsoft-phi-safety-panelsLower is better0.0361%Misuse resistance 0.361%
jailbreakphi-system-card/phi-system-card.csv:jailbreakMeasures jailbreak defect rate.ordinary_harm_misuse_resistance:1.000microsoft-phi-safety-panelsLower is better0.0433%Misuse resistance 0.433%
third_party_harmphi-system-card/phi-system-card.csv:third_party_harmMeasures third-party harm defect rate.human_rights_systemic_harm:1.000microsoft-phi-safety-panelsLower is better0.221%Human rights 1.48%

harmful_continuation

Measures harmful-content continuation defect rate.

RankModelValueRelative performanceProvenance
1phi-3-small0.003official
2phi-3.5-moe-instruct0.004official
3phi-3.5-mini-instruct0.006official
4phi-3-mini0.007official
5phi-3-medium0.01official
6gemma-7b-it0.013official
6llama-3-8b-instruct0.013official
8mistral-nemo0.025official
9mistral-7b-instruct0.026official
10phi-20.029official

harmful_summarization

Measures harmful-content summarization defect rate.

RankModelValueRelative performanceProvenance
1llama-3-8b-instruct0.082official
2phi-3-mini0.1official
3gemma-7b-it0.103official
4mistral-nemo0.11official
4phi-3-small0.11official
6phi-3-medium0.112official
7phi-3.5-mini-instruct0.119official
8phi-3.5-moe-instruct0.12official
9phi-20.144official
10mistral-7b-instruct0.223official

jailbreak

Measures jailbreak defect rate.

RankModelValueRelative performanceProvenance
1phi-3-small0.107official
2phi-3.5-moe-instruct0.11official
3phi-3-medium0.111official
4gemma-7b-it0.114official
5phi-3.5-mini-instruct0.119official
6phi-3-mini0.123official
7llama-3-8b-instruct0.13official
8phi-20.15official
9mistral-7b-instruct0.156official
10mistral-nemo0.174official

third_party_harm

Measures third-party harm defect rate.

RankModelValueRelative performanceProvenance
1phi-3.5-moe-instruct0.152official
2phi-20.24official
2phi-3-mini0.24official
4phi-3.5-mini-instruct0.243official
5phi-3-medium0.251official
6phi-3-small0.253official
7llama-3-8b-instruct0.373official
8gemma-7b-it0.383official
9mistral-7b-instruct0.562official
10mistral-nemo0.599official