Model profile
Gemma 3 4B
Evidence summary
Safety. Gemma 3 4B has an estimated Safety rank of #298; its 90% source-sensitivity interval is #207–#324. Its behavior-only rank is #312; company governance moves the combined estimate to #298. Published Safety evidence spans 6 eval lineages and 6 of 7 components. Its strongest relative result is IndoBias-Pairs — parity-aware culturally grounded bias (parity_score, #8 of 26); its weakest is AA-Omniscience (hallucination_rate, #358 of 358).
Freedom. Gemma 3 4B has an estimated Freedom rank of #57; its 90% source-sensitivity interval is #49–#312. Published Freedom evidence spans 3 eval lineages and 1 of 1 components. Its strongest relative result is ThaiSafetyBench (safety_score, #3 of 18); its weakest is Enkrypt AI Safety Leaderboard (cbrn_attack_non_success_rate, #153 of 270).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gemma-3-4b
- Official model documentation ↗Family-level model document · google · first party
- Release source ↗curated outward evidence
Safety evals
Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #358 / 358 | ↓0.9821 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #75 / 104 | ↓59.02 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #136 / 270 | ↑16.02 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #114 / 270 | ↑88.5 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #210 / 270 | ↑48.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #177 / 268 | ↑95.05 | Source ↗official | |
| IndoBias-Pairs — parity-aware culturally grounded biasparity_score | #8 / 26 | ↑87.05 | Source ↗official | |
| KIDBench Implicit Child Cueimplicit_child_cue_total_mean | #11 / 13 | ↑3.52 | Source ↗official | |
| ThaiSafetyBenchsafety_score | #16 / 18 | ↑71.89 | Source ↗official |
Freedom evals
Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #30 / 104 | ↑59.02 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #153 / 270 | ↓88.5 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #60 / 270 | ↓48.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #91 / 268 | ↓95.05 | Source ↗official | |
| ThaiSafetyBenchsafety_score | #3 / 18 | ↓71.89 | Source ↗official |