Description
Documentation of data sources, collection methods, transformations, and lineage throughout the AI system lifecycle.
Implementation Guidance
Data Lineage Tracking
Implement end-to-end data lineage from source to model: document original data sources (APIs, databases, scraping, purchases), collection methods and consent basis, all transformations applied (cleaning, augmentation, feature engineering), and which model versions used which data versions.
Data Sheets
Create data sheets (per Gebru et al.) for all significant training datasets documenting: motivation, composition, collection process, preprocessing, uses, distribution, and maintenance. Update data sheets when datasets change.
Traceability Tools
Use data versioning tools (DVC, Delta Lake, or equivalent) to maintain immutable snapshots of training data. Enable any production model prediction to be traced back to its exact training data version.
Evidence Requirements
- Data lineage diagrams per AI system
- Data sheets for training datasets
- Data version control logs
- Source documentation and consent records