Benchmarks
Data Canon brings safety benchmarks and public model scores into one place. Every score links back to a source, and missing scores are shown as missing rather than guessed.
Featured benchmarks
6Curated for Phase 1 safety coverageModels covered
3Only where public scores are availableEvidence standard
CitedEvery score links to a sourceLeaderboard
Data Canon only displays model scores with public sources. Missing cells mean no comparable public score was found during the latest review.
| Model | Provider | AILuminate | HarmBench | WMDP | CyberSecEval | TruthfulQA | BBQ |
|---|---|---|---|---|---|---|---|
| Llama 2 70B | Meta | No public score | No public score | No public score | No public score | 63.52%Reported | No public score |
| Llama 2 13B | Meta | No public score | No public score | No public score | No public score | 62.18%Reported | No public score |
| Llama 2 7B | Meta | No public score | No public score | No public score | No public score | 57.04%Reported | No public score |
AILuminate
MLCommons' safety evaluation framework that grades model responses across structured hazard categories including violent crime, chemical/biological weapons, and child safety content. Higher scores indicate stronger safety alignment.
HarmBench
A standardized evaluation framework for red teaming language models, measuring attack success rates across direct requests and automated jailbreak methods. Lower attack success rates indicate stronger resistance to adversarial prompts.
WMDP
A multiple-choice benchmark measuring models' retained hazardous knowledge about biological, chemical, and cyber weapons. Lower accuracy indicates less hazardous knowledge — ideally near the 25% random baseline.
CyberSecEval
Meta's cybersecurity evaluation suite measuring models' propensity to assist with cyberattacks, generate insecure code, and support offensive security operations. Lower scores indicate lower cyber risk.
TruthfulQA
A benchmark measuring whether language models generate truthful answers to 817 questions that probe common misconceptions. Higher scores indicate greater truthfulness and resistance to confidently stating falsehoods.
BBQ
A benchmark measuring social bias in language model outputs across 11 demographic categories including age, race, gender, and religion. Higher accuracy on disambiguated questions indicates less reliance on social stereotypes.