Model and Data Cards
Overview
Section titled “Overview”- Primary references: Mitchell et al., Model Cards for Model Reporting (free); Gebru et al., Datasheets for Datasets (free)
- Supplementary: Hugging Face, model card guide (free); Pushkarna et al., Data Cards (free); the course ships
MODEL_CARD.mdandDATASHEET.mdtemplates with the course contracts when this module lands - Prerequisites: Data Licensing and Privacy and PII; for the course module, the trained capstone model (
C1) - Estimated time: 3 to 4 h in course Pass 9
Key Takeaways
Section titled “Key Takeaways”- A model card is the model’s interface documentation: intended use, out-of-scope use, training data, evaluation results with uncertainty, and known limitations.
- A datasheet answers why the dataset exists, how it was collected and cleaned, what it contains, and what it should not be used for.
- Cards are generated where possible and checked always. The datasheet comes from the ledger (
data.08); the release workflow refuses a model without a card (dur.12). - Third-party models are labelled as such: the agent model you fetch (SmolLM2-135M-Instruct) gets its own provenance line, not yours.
How to Study
Section titled “How to Study”Read both papers and then three real model cards on the Hugging Face Hub, noting what each omits. In the course, ethics.03 writes the datasheet and model card for your capstone model and wires them into the release gate.
Concepts & Techniques
Section titled “Concepts & Techniques”Core Insight
Section titled “Core Insight”Documentation is how a model’s limits travel with it. Without a card, every downstream user re-learns the model’s failure modes in production; with one, the limits are stated, measured, and versioned next to the weights.
1. What a card contains
Section titled “1. What a card contains”Key ideas:
- Model card: details, intended use, factors, metrics, evaluation data, training data, quantitative analyses, ethical considerations, caveats.
- Datasheet: motivation, composition, collection, preprocessing, uses, distribution, maintenance.
2. Cards in the release path
Section titled “2. Cards in the release path”Key ideas:
- Generated sections: corpus statistics and license summary from the ledger, eval tables from the
EvalSuitereport. - Gate:
ModelRelease(dur.12) checks the card exists and its eval section matches the release candidate.
Course modules
Section titled “Course modules”| Module | Topic | Kind | Pass |
|---|---|---|---|
ethics.03 | Datasheet and model card | practice | 9 |
Chapters
Section titled “Chapters”| # | Module | Chapter | Kind | Pass |
|---|---|---|---|---|
| 1 | ethics.03 | Datasheet and model card | practice | 9 |
Connections to Other Tracks
Section titled “Connections to Other Tracks”| Track | Connection |
|---|---|
| Responsible AI | the track overview and how the six topics connect |
| Bias and Safety Evals | the numbers the card reports |
| Documentation Writing | cards as reference documents |
| tinyllm Capstones | the model the card describes |