Writing a model card
A card is the first thing people read. Here is what to put in it.
A good card answers five questions in plain words. Keep each part short; link out for detail.
- What is it? One sentence on the task and the size.
- How was it trained? Data, recipe and compute, honestly.
- How do I use it? A runnable snippet.
- How good is it? Evaluations you actually ran, with the setup.
- Where does it fail? Known gaps, biases and inputs it was never shown.
Metadata
The header of the card carries the task, license, languages and base model. Filters on the models page read those fields, so fill them in.
---
task: chat
license: apache-2.0
languages: [en]
base: kestrel/kestrel-1.5b-base
---