Model profile
Qwen2.5 72B Instruct
Evidence summary
Qwen2.5 72B Instruct has an estimated overall rank of #114; its 90% source-sensitivity interval is #19–#217. Its behavior-only rank is #106; company governance moves the combined estimate to #114. Published evidence spans 16 evals and 7 of 7 behavior components. Its strongest relative result is Enkrypt AI Safety Leaderboard (toxicity_attack_non_success_rate, #1 of 258); its weakest is VETO Misfired Alignment (misfired_alignment_rate_pct, #22 of 25).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗qwen2-5-72b-instruct
- Hugging Face ↗Qwen/Qwen2.5-72B-Instruct
- OpenRouter ↗qwen/qwen-2.5-72b-instruct
- Model card ↗Exact model document · Qwen · first party
- Release source ↗repository created at proxy
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 |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #195 / 311 | 0.8469 | ↓ lower | Source ↗official | |
| Agent-SafetyBenchcompromise_availability | #5 / 16 | 35.2 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchharmful_vulnerable_code | #10 / 16 | 29.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchleak_sensitive_information | #8 / 16 | 32.8 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchphysical_harm | #9 / 16 | 29.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchproduce_unsafe_information | #6 / 16 | 97.2 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchproperty_loss | #8 / 16 | 38 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchspread_unsafe_information | #9 / 16 | 12 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchviolate_law_ethics | #10 / 16 | 24 | ↑ higher | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #51 / 80 | 0.59 | ↑ higher | Source ↗official | |
| Confabulationsconfabulation_rate | #39 / 52 | 32.18 | ↓ lower | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #83 / 260 | 21.45 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #136 / 260 | 88.17 | ↑ 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 | #1 / 258 | 100 | ↑ higher | Source ↗official | |
| FORTRESSaverage_risk_score | #42 / 49 | 56.43 | ↓ lower | Source ↗official | |
| FORTRESSover_refusal_score | #5 / 46 | 1.19 | ↓ lower | Source ↗official | |
| HELM Safetyanthropic_red_team | #13 / 80 | 0.996 | ↑ higher | Source ↗official | |
| HELM Safetybbq | #23 / 80 | 0.954 | ↑ higher | Source ↗official | |
| HELM Safetyharmbench | #42 / 80 | 0.728 | ↑ higher | Source ↗official | |
| HELM Safetysimple_safety_tests | #1 / 80 | 1 | ↑ higher | Source ↗official | |
| HELM Safetyxstest | #11 / 80 | 0.979 | ↑ higher | Source ↗official | |
| PandaBench JBB direct-request panelsafety_rate | #14 / 46 | 0.99 | ↑ higher | Source ↗official | |
| SOSBenchbiology_pvr | #15 / 23 | 0.68 | ↓ lower | Source ↗official | |
| SOSBenchchemistry_pvr | #15 / 23 | 0.56 | ↓ lower | Source ↗official | |
| SOSBenchmedicine_pvr | #15 / 23 | 0.734 | ↓ lower | Source ↗official | |
| SOSBenchpharmacology_pvr | #16 / 23 | 0.926 | ↓ lower | Source ↗official | |
| SOSBenchphysics_pvr | #15 / 23 | 0.678 | ↓ lower | Source ↗official | |
| SOSBenchpsychology_pvr | #16 / 23 | 0.734 | ↓ lower | Source ↗official | |
| SYCON Benchfalse_presupposition_tof | #6 / 11 | 2.35 | ↑ higher | Source ↗official | |
| SYCON Benchunethical_queries_tof | #7 / 11 | 1.32 | ↑ higher | Source ↗official | |
| TACbase_welfare_rate | #4 / 68 | 49.4 | ↑ higher | Source ↗self run | |
| VETO Misfired Alignmentmisfired_alignment_rate_pct | #22 / 25 | 11.4 | ↓ lower | 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.