Model profile
Phi 3 Mini 4K Instruct
Evidence summary
Phi 3 Mini 4K Instruct has an estimated overall rank of #117; its 90% source-sensitivity interval is #21–#245. Its behavior-only rank is #116; company governance moves the combined estimate to #117. Published evidence spans 4 evals and 3 of 7 behavior components. Its strongest relative result is Open LLM Safety Index (jailbreakbench_safety_rate, #1 of 21); its weakest is Enkrypt AI Safety Leaderboard (bias_attack_non_success_rate, #172 of 260).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Hugging Face ↗microsoft/Phi-3-mini-4k-instruct
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- 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 |
|---|---|---|---|---|---|
| DSPSafeBenchscore | #2 / 12 | 78.62 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #172 / 260 | 13.18 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #109 / 260 | 89.5 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #111 / 260 | 75.56 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #144 / 258 | 95.77 | ↑ higher | Source ↗official | |
| Open LLM Safety Indexjailbreakbench_safety_rate | #1 / 21 | 0.8 | ↑ higher | Source ↗official | |
| Open LLM Safety Indexstrongreject_safety_rate | #7 / 21 | 0.5333 | ↑ higher | Source ↗official | |
| PandaBench JBB direct-request panelsafety_rate | #28 / 46 | 0.97 | ↑ 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.