34. Making model cards, Chris Albon, Wikimedia

150 models in prod, language translation, topic detection
Governed by the community and open funded by users, so pretty careful where money is spent on
Widely transparent: all code is public, ML team internal chat is public, tickets are public

No black boxes, no secret sauce, have participation in it. We need a way for users to read and understand the models
Model cards are a single source of truth for an ML model

public-facing, discuss motivation, training, its aim, how to get the code, the data etc

important for transparency
Spent a lot of time speaking to the community, researchers, what do they looking for in a tool like this

If people on french Wikipedia don’t want to use a tool, they can turn it off.
POC was pretty straightforward
- Language agnostic topic detection
- Talk about how the model should be used, all the story behind this. The least technical part of the discussion at the intro.
One model card
Data card