learning-it by example

#213 Data governance 5.5 min

Govern AI Training Data: Provenance, Consent, Dataset Card

Ten records, five used. A dataset card says why the other five were left out.

1. What does a dataset card record, compared with the longer datasheets for datasets?
2. In the Python training data curation demo with ten made-up records, how many records were used, how many were left out, and why were two of them left out for no consent?
3. In the Python training data curation demo, how does the decision code treat a record whose consent value is unknown?
4. In the Python training data curation demo, a record fails two checks, for example it has no licence and it also holds personal data without consent. What does the decision log show?
5. On the dataset card made by the Python training data curation demo, the fingerprint is the start of a hash of what, and what does that mean for edited records?

Practice only: nothing is saved and no score counts towards anything. Each answer comes with the reason.

  1. Record consent for record 6

    Copy the code folder first. In records.py, change the consent value of record 6 from None to True, then run python run.py. Predict first how many records are used and whether the fingerprint changes. When I ran it, 6 were used, 4 were left out, the reasons showed no consent 1, and the fingerprint started 131ee1e4e612. Undo this change afterwards, so the next exercises give the numbers they state.

    A hint

    Record 6 holds personal data. Unknown consent counts as no consent, but True is accepted.

  2. Allow the partner licence

    In curate.py, add "partner-only" to the ALLOWED set, then run python run.py. Predict which record changes and what the Licences line gains. When I ran it, 6 records were used, the card listed partner-only 1 under Licences, and licence not allowed disappeared from the reasons. Undo this change afterwards, so the next exercises give the numbers they state.

    A hint

    Record 5 is the only one with the licence partner-only.

  3. Give record 4 an opt out

    At the end of records.py, add the line RECORDS[3]["opt_out"] = True. Record 4 has no licence. Predict which reason the log shows for record 4 now, and whether the fingerprint changes. When I ran it, record 4 showed opt-out, the reasons were licence not allowed 1, no consent 2, opt-out 2, and the fingerprint stayed 6f31a401470b.

    A hint

    Look at the order of the if statements in decide, and at which records go into the hash.

  4. Change the order of the checks

    In curate.py, move the no licence check above the opt-out check, with record 4 still marked as an opt out from the previous task. Predict what the log shows for record 4, then run it and compare. Write one sentence on why the order of checks changes the reasons you report.

    A hint

    The first check that fails is the only reason returned.

These tasks are for you to try on your own machine. Nothing is checked or saved.

What this does not show

Ten made-up records and my own rules; whether a licence or consent covers training is a legal question, the fingerprint covers record ids only, and this is not legal advice.

Up next in Data governance

More in Data governance

All Data governance videos

Related videos