Engineering, production, governance, cost, communication · insight-0281
Write a datasheet for any dataset other teams will train on
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Data analyst, Data scientist, ML engineer
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A curated training table is published internally, and three teams start building models on it.
Action
Publish a short datasheet with motivation, collection process, time range, inclusion and exclusion rules, labelling method, known gaps and biases, recommended and discouraged uses, owner and update cadence.
Reason
Documenting dataset motivation, composition, collection and recommended uses helps consumers judge fitness for their purpose.
Boundary
Shared feature tables, labelled datasets, evaluation sets, vendor data.
Symptom or risk
Consumers do not know how rows were collected, filtered or labelled, so they reuse the data for purposes it cannot support and inherit its biases unknowingly.
Exceptions
Scratch datasets for one analysis need only a README line, unless they are later shared.
Trade-offs
Datasheets go stale; tie updates to schema or pipeline changes.
Pitfalls
Listing columns without explaining the selection process that created the rows.
Lessons and concepts
Claims and their sources
Gebru and colleagues propose datasheets documenting a dataset's motivation, composition, collection process and recommended uses.