For judgesFor judges
Data engineering
How the open datasets are built, validated, versioned and verified. Prepared, not yet published.
The open datasets (see Open data, prepared, not yet published) come out of a checked pipeline, not a one-off export. Everything below runs in GitHub Actions on every change to the build and on each manual run of the hf-datasets workflow.
What is checked
| Area | Check |
|---|---|
| Schema | Every table has a schema (types, nulls, ranges, enums, key); wrong or missing fields fail the build |
| Money | Integer cents, never negative |
| Ledger | Held equals captured plus released plus open, per lot and per hold, in the simulated run |
| Time | Timestamps and sequence numbers never go backwards; percentiles are ordered |
| Consistency | Scores recompute from outputs; evidence hashes equal the ones in docs/GATES.md; counts add up |
| Privacy | No emails, tokens, key-like strings, IPs or payment-object-like ids; demo actors anonymised |
Reproducible and verifiable
- One command,
scripts/hf/build.sh, runs the chain. The workflow runs it twice and fails if the twoMANIFEST.jsonfiles differ. - Every dataset carries
MANIFEST.json(version, repository commit, build run URL, sha256 and row count of every file, source files) andverify.py, which recomputes the hashes. - Every dataset has a version and a
CHANGELOG.md; the data dictionary and a data-quality report are generated in CI and are in the run artifact.
Honest labels
Real sandbox (PayPal sandbox, small samples, never production), simulated (fake clock, a model of PayPal), synthetic audio (machine-generated speech), fake PayPal (scripted demo auctions). The pipeline shows the data is well formed and traceable; it does not turn a simulation into a measurement. The full write-up is docs/DATA-ENGINEERING.md in the repository.
Edit on GitHub
Last updated on