FundamentalsUpdated 2026-10-24
Model Evaluation
Designing test sets and scoring rules to decide whether a model works for a specific task.
#评测#质量#上线
In plain language
Evaluation should cover normal, edge and adversarial cases across factuality, completeness, format and safety. Human review needs explicit rubrics; automated evaluation can combine rules, model judges and sampling. Monitoring continues after launch.
Example
Build a rubric for 200 real support questions and sample 50 weekly.
Recommended tools
Related terms
Fundamentals
Benchmark
A standardized dataset and metric for comparing model capabilities.
3 related terms
Safety & Alignment
Guardrails
Rules and systems that constrain model inputs, outputs and actions.
3 related terms
FundamentalsFeatured
Hallucination
When a model produces plausible but factually wrong or unsupported content.
3 related terms
Applications
Agentic Workflow
Breaking a task into steps planned by AI, executed by tools and reviewed by humans.
3 related terms