Responsible AI guidelines

What we ask of anyone who publishes data or a model here, and of anyone who uses one.

Before you publish data

Say where it came from and how it was gathered. Provenance is not a formality: it is what lets somebody else judge whether your data fits their problem.
Remove or mask personal details unless you have a lawful basis to publish them and, where consent is the basis, that consent.
Say what is missing. Which districts, which dialects, which kinds of people are not in this data — every dataset is partial, and the ones that admit it are the useful ones.
Say how it was labelled, by whom, and what they were told to do.

Say what it is for, and what it is not for.
Say what it was trained on, and under what licence.
Report how it does — and where it does worst. A model that works on formal Bangla and fails on a dialect must say so.
Say what it costs to run, so somebody with one machine knows whether it is for them.

A model’s answer is a guess, however confident it sounds. Where a wrong answer would hurt somebody — health, money, justice, safety — a person decides, not the model.
Test it on your own people’s data before you rely on it. Performance reported here was measured somewhere else.
Tell the people affected that a machine is involved, and give them a way to reach a human.

Bangla is written by people whose dialect, spelling and script habits differ. Treat variation as the language, not as noise to be cleaned away.
Bengalis live in many countries. Data from one of them does not speak for the rest.
Publish evaluation in Bangla too, so the people a model is about can read how well it works on them.

Review asks for these things, and a curator may return work that lacks them.
A dataset that carries personal data without a basis is taken down.
Work that is used to target or harm a group is removed, and the account that published it is suspended.