All learning paths Learn See Try Make Release it with a datasheet What this lesson is about Publish it with a sheet: what it was made for, who is in it, what is missing. Name the one limit too.
Your partner wants the Santali folder tonight and says the notes can follow next week. What goes out?
A Send the folder tonight, and the written answers next week when there is time for them. B Send the half you are sure about, and hold the rest until the notes are ready. C Answer the datasheet questions in writing: purpose, people, money, gaps and limits.
Check my recall Why this matters Datasheets for Datasets asks a fixed list of questions of every dataset — the pile of examples a model learns from. It expects them answered in writing before the data goes out. Why it was made, who is in it, who paid, what is missing, what it may not be used for. The government's own repository asks the same way. An upload carries its description and its access settings. It is published only after an organisation checker and a platform checker have seen it.
Release a collection with the sheet that answers what a trainer will ask of it.
Datasheets for Datasets asks a fixed list of questions of every dataset — the pile of examples a model learns from. It expects them answered in writing before the data goes out. Why it was made, who is in it, who paid, what is missing, what it may not be used for. The government's own repository asks the same way. An upload carries its description and its access settings. It is published only after an organisation checker and a platform checker have seen it.
Answer the datasheet questions in writing: purpose, people, money, gaps and limits. Choose the terms of use before you upload, because a released set cannot be called back. State counted numbers, not adjectives: hours, speakers, places, and how much is written out. Task Two hundred hours of Maithili speech are ready, and a partner asks for the download link.
Weak approach Sending the folder with a single line beside it. The line says two hundred hours of Maithili speech, collected this year by our team.
Stronger approach Sending it with a sheet. Why it was collected and who paid for it. 212 hours from 340 speakers in four blocks, 61 of those hours written out. Nobody under eighteen. One line saying it may not be used to recognise a speaker, and the licence named on the same page.
Why the stronger approach works Both are the same audio. Only one can be used by somebody who has to say in public where their training material came from.
Try a changed situation A year later a company asks whether your released set can be used to build a voice lock. What decides it?
Make something yourself Write the release sheet for a collection you made or checked. Answer the datasheet questions and name the licence you would choose. State the counted numbers a trainer would ask you for.
Lesson 12 of 12 on Bharat-language data craft next What people actually come with Part 4 ends in a work sample you can send The rules this work sample is read against
Read the primary or official source: Gebru et al., Datasheets for Datasets (arXiv)