FundamentalsUpdated 2026-10-24
Benchmark
A standardized dataset and metric for comparing model capabilities.
#评测#模型#数据
In plain language
Common benchmarks include MMLU, GSM8K, SWE-bench and LMArena. Scores are affected by prompts, harnesses and data contamination, so they do not equal real-world performance. Evaluate on your own task set.
Example
A model can top SWE-bench yet underperform another on your private codebase.
Recommended tools
Related terms
FundamentalsFeatured
Hallucination
When a model produces plausible but factually wrong or unsupported content.
3 related terms
FundamentalsFeatured
Large Language Model
A model trained on massive text to understand and generate natural language.
3 related terms
Fundamentals
Model Evaluation
Designing test sets and scoring rules to decide whether a model works for a specific task.
3 related terms
Applications
Agentic Workflow
Breaking a task into steps planned by AI, executed by tools and reviewed by humans.
3 related terms