Documentation is a maintained collection of information describing a system, its implementation, operation, and procedures. In machine learning systems, it transfers knowledge, supports onboarding, and helps people outside the responsible team understand and run the system. section 16.3
Treat documentation as operational infrastructure rather than deferred administration:
- Keep it available and updated as assumptions, dependencies, sources, and responsibilities change. section 16.3 section 16.6
- Use it to smooth onboarding and make accountability explicit across teams. section 16.3 section 16.6
- Prefer reproducible instructions that let people inspect relevant code and datasets or run deployment with a single command. section 16.3
- Invest early: neglected documentation can turn a two-week reconstruction into a three-month effort. section 16.3 section 16.6
Reproducing a deployed system
A documented language-model project could be reproduced from a specific tag, including its deployment process, code repository, and datasets, without relying on its original builder. section 16.3
