Evaluation Landscape

Explore how benchmarks relate across different lenses - how their results correlate, which domains they intend to cover, and how difficult their items are.

Safety & Alignment

15 benchmarks

• AdvBench
• Anthropic Red Teaming
• BBQ
• BOLD
• Disinformation
• Do-Not-Answer
• HarmBench
• Health-ORSC-Bench
• MoralBench
• OR-Bench
• SALAD-Bench
• SimpleSafetyTests
• Sorry-Bench
• TruthfulQA
• XSTest

Interactive & Real-World Behavior

13 benchmarks

• AdvBench
• Anthropic Red Teaming
• Disinformation
• Do-Not-Answer
• HarmBench
• Health-ORSC-Bench
• IFEval
• OR-Bench
• SALAD-Bench
• SimpleSafetyTests
• Sorry-Bench
• WildBench
• XSTest

Knowledge & Reasoning

9 benchmarks

• BoolQ
• CulturalBench
• DocHop
• GPQA
• Hi-ToM
• MMLU-Pro
• Omni-MATH
• OpenToM
• VisFactor

Social & Human-Centered Reasoning

7 benchmarks

• BBQ
• BOLD
• CulturalBench
• EmoBench
• Hi-ToM
• MoralBench
• OpenToM

Language & Instruction Following

6 benchmarks

• CNN/DailyMail
• IFEval
• IMDb
• Last-Translation-Benchmark
• WildBench
• XSum

Truthfulness & Information Integrity

2 benchmarks

• Disinformation
• TruthfulQA

*Curated benchmark domains. A benchmark may appear in more than one category.