Model profile
Phi 3.5 Moe Instruct
Evidence summary
Phi 3.5 Moe Instruct has an estimated overall rank of #31; its 90% source-sensitivity interval is #8–#158. Its behavior-only rank is #25; company governance moves the combined estimate to #31. Published evidence spans 5 evals and 6 of 7 behavior components. Its strongest relative result is PandaBench JBB direct-request panel (safety_rate, #1 of 46); its weakest is Microsoft Phi Safety Panels (harmful_summarization, #8 of 10).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Hugging Face ↗microsoft/Phi-3.5-MoE-instruct
- Model card ↗Exact model document · Microsoft · first party
- Release source ↗direct research preview identity date
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Better | Source |
|---|---|---|---|---|---|
| AILuminate General Purpose AI Chatchild_sexual_exploitation | #5 / 32 | 8.2 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatdefamation | #5 / 32 | 3.9 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chathate | #5 / 32 | 2.5 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatindiscriminate_weapons_cbrne | #4 / 32 | 6.7 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatintellectual_property | #6 / 32 | 4.5 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatnon_violent_crimes | #3 / 32 | 4.5 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatprivacy | #5 / 32 | 4.5 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatsex_related_crimes | #3 / 32 | 3.7 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatsexual_content | #4 / 32 | 4.7 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatspecialized_advice | #4 / 32 | 6.9 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatsuicide_self_harm | #4 / 32 | 3.7 | ↓ lower | Source ↗official | |
| AILuminate General Purpose AI Chatviolent_crimes | #3 / 32 | 5 | ↓ lower | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #141 / 260 | 14.99 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #148 / 260 | 87.33 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #107 / 260 | 76.11 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #66 / 258 | 98.32 | ↑ higher | Source ↗official | |
| Large-scale Moral Machine experiment on LLMshuman_choice_distance | #16 / 39 | 0.9383 | ↓ lower | Source ↗official | |
| Microsoft Phi Safety Panelsharmful_continuation | #2 / 10 | 0.004 | ↓ lower | Source ↗official | |
| Microsoft Phi Safety Panelsharmful_summarization | #8 / 10 | 0.12 | ↓ lower | Source ↗official | |
| Microsoft Phi Safety Panelsjailbreak | #2 / 10 | 0.11 | ↓ lower | Source ↗official | |
| Microsoft Phi Safety Panelsthird_party_harm | #1 / 10 | 0.152 | ↓ lower | Source ↗official | |
| PandaBench JBB direct-request panelsafety_rate | #1 / 46 | 1 | ↑ higher | Source ↗official |
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.