Benchmark datasets provide standard test sets for comparing model performance across the research community. Purpose: Enable fair comparison, track progress, identify strengths/weaknesses. Major NLP benchmarks: GLUE: 9 language understanding tasks (sentiment, similarity, NLI). SuperGLUE: Harder successor to GLUE. MMLU: 57 subjects testing world knowledge. HellaSwag: Commonsense reasoning. WinoGrande: Coreference resolution. ARC: Science reasoning. TruthfulQA: Factuality. Code benchmarks: HumanEval, MBPP, MultiPL-E, SWE-bench. Reasoning: GSM8K (math), MATH, Big-Bench, BBH. Leaderboards: Papers With Code, HELM, OpenLLM Leaderboard track rankings. Limitations: Benchmark saturation (models overfit), gaming metrics, may not reflect real-world performance, contamination concerns. Best practices: Use multiple benchmarks, include held-out test sets, validate with human evaluation. Creating benchmarks: Need diversity, clear metrics, held-out test sets, regular updates. Current trends: Moving toward harder benchmarks, agentic tasks, real-world problems as older benchmarks become saturated.