Model profile
Qwen 2.5 7B Instruct
Evidence summary
Qwen 2.5 7B Instruct has an estimated overall rank of #198; its 90% source-sensitivity interval is #75–#247. Its behavior-only rank is #189; company governance moves the combined estimate to #198. Published evidence spans 16 evals and 7 of 7 behavior components. Its strongest relative result is TAC (base_welfare_rate, #6 of 68); its weakest is Agent-SafetyBench (harmful_vulnerable_code, #16 of 16).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Hugging Face ↗Qwen/Qwen2.5-7B-Instruct
- OpenRouter ↗qwen/qwen-2.5-7b-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 |
|---|---|---|---|---|---|
| Agent-SafetyBenchcompromise_availability | #15 / 16 | 17.2 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchharmful_vulnerable_code | #16 / 16 | 10.8 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchleak_sensitive_information | #15 / 16 | 13.2 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchphysical_harm | #15 / 16 | 17.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchproduce_unsafe_information | #15 / 16 | 57.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchproperty_loss | #15 / 16 | 15.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchspread_unsafe_information | #13 / 16 | 7.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchviolate_law_ethics | #15 / 16 | 10.4 | ↑ higher | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #67 / 80 | 0.47 | ↑ higher | Source ↗official | |
| DSPSafeBenchscore | #5 / 12 | 73.51 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #81 / 260 | 21.71 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #201 / 260 | 82.67 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #215 / 260 | 42.22 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #252 / 258 | 61.82 | ↑ higher | Source ↗official | |
| HELM Safetyanthropic_red_team | #49 / 80 | 0.985 | ↑ higher | Source ↗official | |
| HELM Safetybbq | #54 / 80 | 0.906 | ↑ higher | Source ↗official | |
| HELM Safetyharmbench | #48 / 80 | 0.677 | ↑ higher | Source ↗official | |
| HELM Safetysimple_safety_tests | #65 / 80 | 0.96 | ↑ higher | Source ↗official | |
| HELM Safetyxstest | #30 / 80 | 0.966 | ↑ higher | Source ↗official | |
| M3-SafetyBenchoverall_score | #8 / 19 | 92.37 | ↑ higher | Source ↗official | |
| PandaBench JBB direct-request panelsafety_rate | #14 / 46 | 0.99 | ↑ higher | Source ↗official | |
| SafeDialBenchaggression | #16 / 18 | 7.013 | ↑ higher | Source ↗official | |
| SafeDialBenchethics | #15 / 18 | 7.357 | ↑ higher | Source ↗official | |
| SafeDialBenchfairness | #7 / 18 | 7.553 | ↑ higher | Source ↗official | |
| SafeDialBenchlegality | #15 / 18 | 7.21 | ↑ higher | Source ↗official | |
| SafeDialBenchmorality | #16 / 18 | 7.06 | ↑ higher | Source ↗official | |
| SafeDialBenchprivacy | #18 / 18 | 7.05 | ↑ higher | Source ↗official | |
| Shelleducation_jsr | #12 / 14 | 0.804 | ↓ lower | Source ↗official | |
| Shellfinance_jsr | #14 / 14 | 0.914 | ↓ lower | Source ↗official | |
| Shellmanagement_jsr | #14 / 14 | 0.938 | ↓ lower | Source ↗official | |
| SYCON Benchfalse_presupposition_tof | #8 / 11 | 1.93 | ↑ higher | Source ↗official | |
| SYCON Benchunethical_queries_tof | #11 / 11 | 0.72 | ↑ higher | Source ↗official | |
| TACbase_welfare_rate | #6 / 68 | 46.2 | ↑ higher | Source ↗self run | |
| UAVBench safety-critical decision recognitionethical_safety_critical_accuracy | #26 / 27 | 0.535 | ↑ 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.