Write a model card and datasheet to document AI limits
11 steps · about 5 h 10 min · Intermediate
- Model Cards (article, Hugging Face, 10 min)
A short overview of what model cards are and why the practice exists, so the later steps make sense.
- Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification (article, Buolamwini & Gebru, PMLR, 40 min)
A real case where a single overall accuracy number hid large subgroup error gaps. That gap is what documentation is meant to expose.
- Model Cards for Model Reporting (article, Mitchell et al., arXiv, 45 min)
The original framework. It defines the sections and the disaggregated evaluation that every later template builds on.
- Datasheets for Datasets (article, Gebru et al., arXiv, 50 min)
The dataset counterpart: the questions to answer about how data was collected, what it contains, and what it should not be used for.
- Model Card Guidebook (doc, Hugging Face, 25 min)
Shows how the original idea became a practical template, and who reads a card (developers, policymakers, affected people).
- Building a model card (doc, Hugging Face Course, 15 min)
A short walkthrough of each section, including how to write the 'intended uses & limitations' part.
- Model Cards (doc, Hugging Face Hub docs, 15 min)
Covers the metadata (license, datasets, base model, evaluation results) and the template you will fill in.
- Create and share Model Cards (doc, Hugging Face, 25 min)
Hands-on: build a card in code from a template and push it to the Hub.
- Dataset Cards (doc, Hugging Face Hub docs, 15 min)
Turns the datasheet questions into a working dataset card, with metadata and example cards to read.
- tensorflow/model-card-toolkit (repo, Google / TensorFlow (GitHub), 30 min)
Shows how a card can be generated from code with a JSON schema. The repo is archived, so treat it as a reference.
- model-cards-and-datasheets (repo, ivylee (GitHub), 40 min)
A collection of real model cards, system cards and datasheets to compare against your own draft.