Model profile
ChatGLM3 6B
Z.aideveloper
2023-10-27release date
#273 / 346Safety rank
#161 / 662Freedom rank
Evidence summary
Safety. ChatGLM3 6B has an estimated Safety rank of #273; its 90% source-sensitivity interval is #200–#293. Its behavior-only rank is #275; company governance moves the combined estimate to #273. Published Safety evidence spans 10 eval lineages and 4 of 7 components. Its strongest relative result is SafeDialBench (ethics, #4 of 18); its weakest is ChiSafetyBench (mcq_score, #12 of 12).
Freedom. ChatGLM3 6B has an estimated Freedom rank of #161; its 90% source-sensitivity interval is #148–#284. Published Freedom evidence spans 8 eval lineages and 1 of 1 components. Its strongest relative result is SuperCLUE Safety (traditional_safety, #2 of 31); its weakest is SafeDialBench (ethics, #15 of 18).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Hugging Face ↗zai-org/chatglm3-6b
- 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 |
|---|---|---|---|---|
| ChiSafetyBenchharmful_response_rate | #12 / 14 | ↓1.08 | Source ↗official | |
| ChiSafetyBenchmcq_score | #12 / 12 | ↑41.16 | Source ↗official | |
| CMoralEvalfamilial_morality | #11 / 26 | ↑0.43 | Source ↗official | |
| CMoralEvalinternet_ethics | #12 / 26 | ↑0.4 | Source ↗official | |
| CMoralEvalpersonal_morality | #11 / 26 | ↑0.42 | Source ↗official | |
| CMoralEvalprofessional_ethics | #11 / 26 | ↑0.43 | Source ↗official | |
| CMoralEvalsocial_morality | #11 / 26 | ↑0.43 | Source ↗official | |
| Fake Alignment (FINE)multiple_choice_safe_decision_rate | #9 / 14 | ↑45.33 | Source ↗official | |
| Fake Alignment (FINE)open_ended_safe_response_rate | #9 / 14 | ↑94.67 | Source ↗official | |
| FinEval Financial Security Knowledgefinancial_security_accuracy_pct | #17 / 19 | ↑48.2 | Source ↗official | |
| FLAMESdata_protection | #10 / 13 | ↑38.16 | Source ↗official | |
| FLAMESfairness | #7 / 13 | ↑37.75 | Source ↗official | |
| FLAMESlegality | #12 / 13 | ↑28.26 | Source ↗official | |
| FLAMESmorality | #6 / 13 | ↑44.78 | Source ↗official | |
| FLAMESsafety | #7 / 13 | ↑32.63 | Source ↗official | |
| S-Evalbase_en_overall | #12 / 22 | ↑57.7 | Source ↗official | |
| SafeDialBenchaggression | #15 / 18 | ↑7.017 | Source ↗official | |
| SafeDialBenchethics | #4 / 18 | ↑7.637 | Source ↗official | |
| SafeDialBenchfairness | #13 / 18 | ↑7.187 | Source ↗official | |
| SafeDialBenchlegality | #6 / 18 | ↑7.983 | Source ↗official | |
| SafeDialBenchmorality | #9 / 18 | ↑7.243 | Source ↗official | |
| SafeDialBenchprivacy | #9 / 18 | ↑7.567 | Source ↗official | |
| SALAD-Benchattack_enhanced_human_autonomy_integrity | #18 / 33 | ↑12.72 | Source ↗official | |
| SALAD-Benchattack_enhanced_information_safety_harms | #17 / 33 | ↑12.05 | Source ↗official | |
| SALAD-Benchattack_enhanced_malicious_use | #17 / 33 | ↑11.01 | Source ↗official | |
| SALAD-Benchattack_enhanced_misinformation_harms | #17 / 33 | ↑13.16 | Source ↗official | |
| SALAD-Benchattack_enhanced_representation_toxicity | #17 / 33 | ↑12.57 | Source ↗official | |
| SALAD-Benchattack_enhanced_socioeconomic_harms | #14 / 33 | ↑17.75 | Source ↗official | |
| SALAD-Benchbase_human_autonomy_integrity | #20 / 33 | ↑92.55 | Source ↗official | |
| SALAD-Benchbase_information_safety_harms | #23 / 33 | ↑92.21 | Source ↗official | |
| SALAD-Benchbase_malicious_use | #20 / 33 | ↑91.15 | Source ↗official | |
| SALAD-Benchbase_misinformation_harms | #22 / 33 | ↑91.38 | Source ↗official | |
| SALAD-Benchbase_representation_toxicity | #19 / 33 | ↑88.73 | Source ↗official | |
| SALAD-Benchbase_socioeconomic_harms | #19 / 33 | ↑86.96 | Source ↗official | |
| SALAD-Benchmcq_human_autonomy_integrity | #25 / 33 | ↑17.78 | Source ↗official | |
| SALAD-Benchmcq_information_safety_harms | #21 / 33 | ↑20.56 | Source ↗official | |
| SALAD-Benchmcq_malicious_use | #22 / 33 | ↑19.55 | Source ↗official | |
| SALAD-Benchmcq_misinformation_harms | #22 / 33 | ↑20.95 | Source ↗official | |
| SALAD-Benchmcq_representation_toxicity | #24 / 33 | ↑17.81 | Source ↗official | |
| SALAD-Benchmcq_socioeconomic_harms | #22 / 33 | ↑22.78 | Source ↗official | |
| SORRY-Benchavg | #34 / 51 | ↓0.36 | Source ↗official | |
| SuperCLUE Safetyinstruction_attack | #9 / 31 | ↑68.97 | Source ↗official | |
| SuperCLUE Safetyresponsible_ai | #17 / 31 | ↑56.36 | Source ↗official | |
| SuperCLUE Safetytraditional_safety | #28 / 31 | ↑65.96 | 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 |
|---|---|---|---|---|
| ChiSafetyBenchharmful_response_rate | #2 / 14 | ↑1.08 | Source ↗official | |
| ChiSafetyBenchrefusal_rr1 | #9 / 14 | ↓73.38 | Source ↗official | |
| ChiSafetyBenchrefusal_rr2 | #9 / 14 | ↓73.38 | Source ↗official | |
| Fake Alignment (FINE)open_ended_safe_response_rate | #5 / 14 | ↓94.67 | Source ↗official | |
| FLAMESdata_protection | #4 / 13 | ↓38.16 | Source ↗official | |
| FLAMESfairness | #7 / 13 | ↓37.75 | Source ↗official | |
| FLAMESlegality | #1 / 13 | ↓28.26 | Source ↗official | |
| FLAMESmorality | #7 / 13 | ↓44.78 | Source ↗official | |
| FLAMESsafety | #7 / 13 | ↓32.63 | Source ↗official | |
| S-Evalbase_en_overall | #11 / 22 | ↓57.7 | Source ↗official | |
| SafeDialBenchaggression | #4 / 18 | ↓7.017 | Source ↗official | |
| SafeDialBenchethics | #15 / 18 | ↓7.637 | Source ↗official | |
| SafeDialBenchfairness | #5 / 18 | ↓7.187 | Source ↗official | |
| SafeDialBenchlegality | #13 / 18 | ↓7.983 | Source ↗official | |
| SafeDialBenchmorality | #10 / 18 | ↓7.243 | Source ↗official | |
| SafeDialBenchprivacy | #10 / 18 | ↓7.567 | Source ↗official | |
| SALAD-Benchattack_enhanced_human_autonomy_integrity | #16 / 33 | ↓12.72 | Source ↗official | |
| SALAD-Benchattack_enhanced_information_safety_harms | #17 / 33 | ↓12.05 | Source ↗official | |
| SALAD-Benchattack_enhanced_malicious_use | #17 / 33 | ↓11.01 | Source ↗official | |
| SALAD-Benchattack_enhanced_misinformation_harms | #17 / 33 | ↓13.16 | Source ↗official | |
| SALAD-Benchattack_enhanced_representation_toxicity | #17 / 33 | ↓12.57 | Source ↗official | |
| SALAD-Benchattack_enhanced_socioeconomic_harms | #20 / 33 | ↓17.75 | Source ↗official | |
| SALAD-Benchbase_human_autonomy_integrity | #14 / 33 | ↓92.55 | Source ↗official | |
| SALAD-Benchbase_information_safety_harms | #11 / 33 | ↓92.21 | Source ↗official | |
| SALAD-Benchbase_malicious_use | #14 / 33 | ↓91.15 | Source ↗official | |
| SALAD-Benchbase_misinformation_harms | #12 / 33 | ↓91.38 | Source ↗official | |
| SALAD-Benchbase_representation_toxicity | #15 / 33 | ↓88.73 | Source ↗official | |
| SALAD-Benchbase_socioeconomic_harms | #15 / 33 | ↓86.96 | Source ↗official | |
| SORRY-Benchavg | #15 / 51 | ↑0.36 | Source ↗official | |
| SuperCLUE Safetyinstruction_attack | #18 / 31 | ↓68.97 | Source ↗official | |
| SuperCLUE Safetytraditional_safety | #2 / 31 | ↓65.96 | Source ↗official |