Model profile
Llama 3.1 70B Instruct
Evidence summary
Llama 3.1 70B Instruct has an estimated overall rank of #187; its 90% source-sensitivity interval is #53–#213. Its behavior-only rank is #177; company governance moves the combined estimate to #187. Published evidence spans 19 evals and 7 of 7 behavior components. Its strongest relative result is FORTRESS (over_refusal_score, #3 of 46); its weakest is LLM Ethics Benchmark (score, #5 of 5).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗llama-3-1-instruct-70b
- Hugging Face ↗meta-llama/Meta-Llama-3.1-70B-Instruct
- NVIDIA ↗qc69jvmznzxy/llama-3_1-70b-instruct
- OpenRouter ↗meta-llama/llama-3.1-70b-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 |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #117 / 311 | 0.7472 | ↓ lower | Source ↗official | |
| AbstentionBenchanswer_unknown_f1 | #13 / 20 | 0.8615 | ↑ higher | Source ↗official | |
| AbstentionBenchfalse_premise_f1 | #10 / 20 | 0.676 | ↑ higher | Source ↗official | |
| AbstentionBenchstale_f1 | #14 / 20 | 0.6032 | ↑ higher | Source ↗official | |
| AbstentionBenchsubjective_f1 | #13 / 20 | 0.7183 | ↑ higher | Source ↗official | |
| AbstentionBenchunderspecified_context_f1 | #5 / 20 | 0.6927 | ↑ higher | Source ↗official | |
| AbstentionBenchunderspecified_intent_f1 | #15 / 20 | 0.702 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchcompromise_availability | #12 / 16 | 24 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchharmful_vulnerable_code | #10 / 16 | 29.6 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchleak_sensitive_information | #14 / 16 | 20 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchphysical_harm | #14 / 16 | 23.2 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchproduce_unsafe_information | #9 / 16 | 93.2 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchproperty_loss | #13 / 16 | 28.4 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchspread_unsafe_information | #11 / 16 | 10.8 | ↑ higher | Source ↗official | |
| Agent-SafetyBenchviolate_law_ethics | #12 / 16 | 20.4 | ↑ higher | Source ↗official | |
| AgentHarmharm_score | #5 / 12 | 14 | ↓ lower | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #72 / 80 | 0.425 | ↑ higher | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #55 / 105 | 38.98 | ↓ lower | Source ↗official | |
| Contextual MoralChoicehuman_agreement | #6 / 22 | 0.45 | ↑ higher | Source ↗official | |
| FORTRESSaverage_risk_score | #31 / 49 | 44.18 | ↓ lower | Source ↗official | |
| FORTRESSover_refusal_score | #3 / 46 | 1.09 | ↓ lower | Source ↗official | |
| HELM Safetyanthropic_red_team | #74 / 80 | 0.932 | ↑ higher | Source ↗official | |
| HELM Safetybbq | #23 / 80 | 0.954 | ↑ higher | Source ↗official | |
| HELM Safetyharmbench | #74 / 80 | 0.469 | ↑ higher | Source ↗official | |
| HELM Safetysimple_safety_tests | #72 / 80 | 0.925 | ↑ higher | Source ↗official | |
| HELM Safetyxstest | #54 / 80 | 0.945 | ↑ higher | Source ↗official | |
| Large-scale Moral Machine experiment on LLMshuman_choice_distance | #4 / 39 | 0.7398 | ↓ lower | Source ↗official | |
| LLM Ethics Benchmarkscore | #5 / 5 | 75.8 | ↑ higher | Source ↗official | |
| PandaBench JBB direct-request panelsafety_rate | #35 / 46 | 0.95 | ↑ higher | Source ↗official | |
| PropensityBenchscore | #9 / 14 | 55.4 | ↓ lower | Source ↗official | |
| SORRY-Benchavg | #38 / 51 | 0.39 | ↓ lower | Source ↗official | |
| VETO Misfired Alignmentmisfired_alignment_rate_pct | #14 / 25 | 8.6 | ↓ 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.