Model profile
Qwen1.5 72B Chat
Evidence summary
Safety. Qwen1.5 72B Chat has an estimated Safety rank of #198; its 90% source-sensitivity interval is #117–#234. Its behavior-only rank is #189; company governance moves the combined estimate to #198. Published Safety evidence spans 13 eval lineages and 6 of 7 components. Its strongest relative result is SALAD-Bench (mcq_representation_toxicity, #2 of 33); its weakest is CRiskEval (situational_awareness, #15 of 17).
Freedom. Qwen1.5 72B Chat has an estimated Freedom rank of #297; its 90% source-sensitivity interval is #199–#409. Published Freedom evidence spans 10 eval lineages and 1 of 1 components. Its strongest relative result is AIRBench 2024 Safety Scenarios (safety_scenarios, #15 of 80); its weakest is ChiSafetyBench (refusal_rr1, #11 of 14).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Hugging Face ↗Qwen/Qwen1.5-72B-Chat
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗direct research preview identity date
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 |
|---|---|---|---|---|
| AIRBench 2024 Safety Scenariossafety_scenarios | #66 / 80 | ↑0.486 | Source ↗official | |
| ChineseSafescore | #5 / 22 | ↑63.67 | Source ↗official | |
| ChiSafetyBenchharmful_response_rate | #2 / 14 | ↓0.22 | Source ↗official | |
| ChiSafetyBenchmcq_score | #1 / 12 | ↑91.13 | Source ↗official | |
| COMPL-AI AI-Identity Disclosurescore | #11 / 14 | ↑0.726 | Source ↗official | |
| COMPL-AI LLM RuLES Multi-Turn Rule Followingscore | #7 / 14 | ↑0.4856 | Source ↗official | |
| COMPL-AI TensorTrust Goal-Hijacking Resistancescore | #8 / 13 | ↑0.4536 | Source ↗official | |
| CRiskEvaldeception_willingness | #6 / 17 | ↓20.12 | Source ↗official | |
| CRiskEvaldesire_for_resource | #7 / 17 | ↓31.71 | Source ↗official | |
| CRiskEvalharmful_goal | #4 / 17 | ↓33.39 | Source ↗official | |
| CRiskEvalimprovement_intent | #7 / 17 | ↓48.6 | Source ↗official | |
| CRiskEvalmalicious_coordination | #6 / 17 | ↓8.07 | Source ↗official | |
| CRiskEvalself_preservation | #8 / 17 | ↓36.86 | Source ↗official | |
| CRiskEvalsituational_awareness | #15 / 17 | ↓68.75 | Source ↗official | |
| HELM Safetyanthropic_red_team | #37 / 80 | ↑0.99 | Source ↗official | |
| HELM Safetybbq | #64 / 80 | ↑0.846 | Source ↗official | |
| HELM Safetyharmbench | #56 / 80 | ↑0.648 | Source ↗official | |
| HELM Safetysimple_safety_tests | #32 / 80 | ↑0.99 | Source ↗official | |
| HELM Safetyxstest | #42 / 80 | ↑0.957 | Source ↗official | |
| OR-Benchover_refusal_rate | #12 / 25 | ↓46.9 | Source ↗official | |
| OR-Benchtoxic_acceptance_rate | #16 / 25 | ↓5.6 | Source ↗official | |
| SALAD-Benchattack_enhanced_human_autonomy_integrity | #13 / 33 | ↑20.47 | Source ↗official | |
| SALAD-Benchattack_enhanced_information_safety_harms | #13 / 33 | ↑17.92 | Source ↗official | |
| SALAD-Benchattack_enhanced_malicious_use | #13 / 33 | ↑17.05 | Source ↗official | |
| SALAD-Benchattack_enhanced_misinformation_harms | #14 / 33 | ↑18.42 | Source ↗official | |
| SALAD-Benchattack_enhanced_representation_toxicity | #16 / 33 | ↑14.19 | Source ↗official | |
| SALAD-Benchattack_enhanced_socioeconomic_harms | #15 / 33 | ↑14.29 | Source ↗official | |
| SALAD-Benchbase_human_autonomy_integrity | #14 / 33 | ↑96.1 | Source ↗official | |
| SALAD-Benchbase_information_safety_harms | #10 / 33 | ↑96.89 | Source ↗official | |
| SALAD-Benchbase_malicious_use | #15 / 33 | ↑95.2 | Source ↗official | |
| SALAD-Benchbase_misinformation_harms | #17 / 33 | ↑93.65 | Source ↗official | |
| SALAD-Benchbase_representation_toxicity | #16 / 33 | ↑90.06 | Source ↗official | |
| SALAD-Benchbase_socioeconomic_harms | #12 / 33 | ↑91.89 | Source ↗official | |
| SALAD-Benchmcq_human_autonomy_integrity | #2 / 33 | ↑83.89 | Source ↗official | |
| SALAD-Benchmcq_information_safety_harms | #2 / 33 | ↑80.28 | Source ↗official | |
| SALAD-Benchmcq_malicious_use | #2 / 33 | ↑84.81 | Source ↗official | |
| SALAD-Benchmcq_misinformation_harms | #2 / 33 | ↑80 | Source ↗official | |
| SALAD-Benchmcq_representation_toxicity | #2 / 33 | ↑79.27 | Source ↗official | |
| SALAD-Benchmcq_socioeconomic_harms | #2 / 33 | ↑78.89 | Source ↗official | |
| SORRY-Benchavg | #34 / 51 | ↓0.36 | 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 |
|---|---|---|---|---|
| AIRBench 2024 Safety Scenariossafety_scenarios | #15 / 80 | ↓0.486 | Source ↗official | |
| ChiSafetyBenchharmful_response_rate | #11 / 14 | ↑0.22 | Source ↗official | |
| ChiSafetyBenchrefusal_rr1 | #11 / 14 | ↓73.81 | Source ↗official | |
| ChiSafetyBenchrefusal_rr2 | #11 / 14 | ↓73.81 | Source ↗official | |
| COMPL-AI LLM RuLES Multi-Turn Rule Followingscore | #8 / 14 | ↓0.4856 | Source ↗official | |
| HELM Safetyanthropic_red_team | #42 / 80 | ↓0.99 | Source ↗official | |
| HELM Safetyharmbench | #25 / 80 | ↓0.648 | Source ↗official | |
| HELM Safetysimple_safety_tests | #42 / 80 | ↓0.99 | Source ↗official | |
| HELM Safetyxstest | #42 / 80 | ↑0.957 | Source ↗official | |
| OR-Benchover_refusal_rate | #12 / 25 | ↓46.9 | Source ↗official | |
| OR-Benchtoxic_acceptance_rate | #10 / 25 | ↑5.6 | Source ↗official | |
| SALAD-Benchattack_enhanced_human_autonomy_integrity | #20 / 33 | ↓20.47 | Source ↗official | |
| SALAD-Benchattack_enhanced_information_safety_harms | #21 / 33 | ↓17.92 | Source ↗official | |
| SALAD-Benchattack_enhanced_malicious_use | #21 / 33 | ↓17.05 | Source ↗official | |
| SALAD-Benchattack_enhanced_misinformation_harms | #20 / 33 | ↓18.42 | Source ↗official | |
| SALAD-Benchattack_enhanced_representation_toxicity | #18 / 33 | ↓14.19 | Source ↗official | |
| SALAD-Benchattack_enhanced_socioeconomic_harms | #19 / 33 | ↓14.29 | Source ↗official | |
| SALAD-Benchbase_human_autonomy_integrity | #19 / 33 | ↓96.1 | Source ↗official | |
| SALAD-Benchbase_information_safety_harms | #24 / 33 | ↓96.89 | Source ↗official | |
| SALAD-Benchbase_malicious_use | #19 / 33 | ↓95.2 | Source ↗official | |
| SALAD-Benchbase_misinformation_harms | #17 / 33 | ↓93.65 | Source ↗official | |
| SALAD-Benchbase_representation_toxicity | #18 / 33 | ↓90.06 | Source ↗official | |
| SALAD-Benchbase_socioeconomic_harms | #22 / 33 | ↓91.89 | Source ↗official | |
| SORRY-Benchavg | #15 / 51 | ↑0.36 | Source ↗official |