Machine Learning Foundations - Dataset Policy

Status: Draft Applies to: all course datasets, notebooks, downloads, and examples

Rules

1. Prefer original fictional or synthetic data created for this course.
2. Do not use real personal, confidential, employer, client, health, financial, authentication, legal-case, or security-sensitive data.
3. Record creator, date, version, licence, permitted use, modifications, and checksum.
4. Include a data dictionary, expected schema, missing-value rules, and known limitations.
5. Mark every column as available before prediction time, unavailable after prediction time, target, metadata, or excluded.
6. Keep raw/source data read-only and publish transformed outputs under new names.
7. Do not download live remote data during learner exercises.
8. Do not ask learners to upload or paste private data.
9. Include safe examples of missing values, imbalance, timing, and leakage where the lesson needs them.
10. Remove a dataset from publication if rights, provenance, privacy, or safety cannot be demonstrated.

Synthetic-data standard

Synthetic data must be realistic enough to teach the concept but not traceable to real people. Fictional identifiers must not resemble real names, emails, phone numbers, addresses, or account numbers.

Prediction-time standard

Every ML dataset must identify which columns would be available at the prediction time. Columns created after the prediction time must not be used as features in modelling notebooks.

Review before publication

Before publication, each dataset must pass:

• schema check;
• missing-value check;
• duplicate identifier check;
• secret and personal-data scan;
• target leakage review;
• provenance review; and
• clean notebook rerun.