Model Card Generator

Fill in what you know about a model and get a publishable card, with an evaluation table built from your own measurements and optional front matter for a model host.

Model name

Version

OwnerA team or a person who can answer a question about it.

Contact

LicenseThe one on the weights, and the one on the data if they differ.

Base modelWhat it was fine-tuned from, if anything.

DateWhen this version was produced.

Evaluation resultsOne per line: metric = value on dataset. Lines that do not parse are reported, not dropped. Leave it empty and the card will say no results are reported.

What this model doesOne paragraph. The task, the input and the output.

Intended users and usesWho it was built for, and the situations it was tested in.

Out-of-scope usesThe uses you know it is not fit for. This section is the one readers trust the card for.

Training dataWhere it came from, how much, over what period, and what was excluded.

Known limitationsWhat it gets wrong, and where you have seen it fail.

Bias, fairness and harmsWhat you tested for, on what data, and what you found — including nothing.

Safety mitigationsFilters, refusals, rate limits, human review — whatever is actually in front of this model.

Compute and environmentHardware, training time, and any energy figure you actually measured.

How to cite thisA citation, a DOI, or a link.

Output format

Add YAML front matter to the Markdown, for hosts that read it

Sections answered

4 of 9

A card is read for the sections most people leave out — out-of-scope uses, limitations, what was not measured. Those are printed as holes here rather than filled in for you.

Evaluation results listed: 3
Results with no dataset named: 0
Sections left blank: 5
Words: 275

Your model cardRead-only. Select all and copy.

---
# license: TODO
# base_model: TODO
model_name: orders-classifier
version: 1.2.0
---
# orders-classifier model card

- **Model:** orders-classifier
- **Version:** 1.2.0

## What this model does

Classifies an inbound support message into one of eleven routing categories.

## Intended users and uses

The support tooling team, inside the ticket router. It has only been tested on English-language messages that arrived by email or the web form.

## Out-of-scope uses

Not for deciding whether a refund is owed, and not for anything a customer sees directly — it produces a queue name, not a reply. Not tested on chat transcripts, on voice transcriptions, or on any language other than English.

## Evaluation results

| Metric | Value | Dataset |
| --- | --- | --- |
| macro F1 | 0.87 | support-tickets-holdout-2026-06 |
| accuracy | 0.91 | support-tickets-holdout-2026-06 |
| refusal rate | 0.4% | adversarial-probe-set-v2 |

## Training data

> **Not answered.** Where it came from, how much, over what period, and what was excluded.

## Known limitations

Short messages (under about ten words) route to the general queue far more often than they should. Messages mentioning two products are routed on whichever is named first.

## Bias, fairness and harms

> **Not answered.** What you tested for, on what data, and what you found — including nothing.

## Safety mitigations

> **Not answered.** Filters, refusals, rate limits, human review — whatever is actually in front of this model.

## Compute and environment

> **Not answered.** Hardware, training time, and any energy figure you actually measured.

## How to cite this

> **Not answered.** A citation, a DOI, or a link.

Still to write — these appear as holes in the card above:

Owner
Contact
License
Date
Base model
Training data
Bias, fairness and harms
Safety mitigations
Compute and environment
How to cite this

What this assumes: the content this page opens with is an example for an invented model on invented datasets — replace all of it, including the numbers. Nothing is inferred and nothing is filled in for you. If the evaluation box is empty the card says no results are reported, because a card that implied a model had been measured when it had not would be worse than no card at all. A result with no dataset named is kept and marked “not stated” rather than dropped — the number is still yours, it is just not yet a claim anyone can check. The front matter keys are a common convention on model hosts, not a standard. Everything on this page runs in your browser. Nothing you paste is uploaded, logged or sent anywhere.

A model card is not documentation of how a model works. It is a statement of what its authors are prepared to claim, and the sections that carry the weight are the negative ones: what it should not be used for, what it was not tested on, where it is known to fail. A card full of capabilities and empty of limitations tells a careful reader something specific, and it is not flattering.

A number without a dataset is not a result

“91% accurate” is unfalsifiable. “91% accurate on support-tickets-holdout-2026-06” can be argued with, reproduced, or found to be leaking training data — all of which are better outcomes than being believed. That is why the evaluation box above wants the set named, and why a line without one is kept but marked rather than silently tidied up.

Write it before you ship, not after

The sections that are hard to fill in are hard because the work behind them has not been done, which is exactly what a card is for discovering. If you cannot say what data the model was trained on, you cannot answer a licensing question later. If you cannot name an out-of-scope use, nobody has yet asked what happens when someone points this at a task it was never meant for. Filling in the holes is the work; the document is the receipt.

Model Card Generator

A number without a dataset is not a result

Write it before you ship, not after

The reasoning behind this

Related tools