What is a model card in machine learning?

A model card is a short document that travels with a trained model, describing what it does, who made it, how it was trained and evaluated, and where it should not be used. The idea comes from Mitchell et al. (2019). On the Hugging Face Hub, a repository's README.md is its model card.

What should a Hugging Face model card include?

A YAML metadata block (license, language, library_name, pipeline_tag, datasets, tags, metrics, and an optional model-index), then Markdown sections: Model Details, Uses, Bias/Risks/Limitations, How to Get Started, Training Details, Evaluation, Environmental Impact, and Citation. This tool emits every one of those in the documented order.

What YAML metadata goes in a model card README?

The keys the Hub recognises are license, language, library_name, pipeline_tag, base_model, datasets, tags, metrics, and model-index. They sit between two --- fences at the very top of README.md. This generator emits only the keys you fill in, so the card stays clean and Hub-valid.

What license tag do I put on a Hugging Face model?

Use the identifier from Hugging Face's licenses list, e.g. apache-2.0, mit, cc-by-4.0, cc-by-nc-4.0 (non-commercial), llama3, or gemma. Pick the one that matches your training data and weights. If nothing fits, use other. The license select in this tool lists the common values verbatim.

What is the difference between a model card and a dataset card?

A model card documents a trained model and lives in the model repo's README.md; a dataset card documents a dataset and lives in the dataset repo's README.md. They use different YAML schemas — for example a dataset card uses task_categories and configs, not pipeline_tag or model-index. This tool builds model cards only.

Why does the generated card have [More Information Needed] in it?

That is the exact placeholder Hugging Face's own annotated template uses for sections you have not filled in yet. Leaving it in is intentional: the card stays complete and honestly signals the gaps instead of hiding them. Replace each placeholder as you gather the details.

When does the tool add a model-index block?

Only when you enter at least one evaluation result (a metric, its value and the dataset). Each row becomes one entry under model-index → results, with the task type taken from your pipeline_tag. With zero rows, both the metrics list and model-index are omitted so the YAML stays valid.

Is anything I type uploaded to a server?

No. The whole card is assembled in your browser with string operations — no API keys, no network requests. You can go offline after the page loads and it still works. Copy the Markdown or download README.md and commit it to your Hub repo.

Developer · AI

AI Model Card Generator — Hugging Face README & YAML Metadata

Fill a form and get a complete, Hub-valid README.md model card — correctly-keyed YAML frontmatter plus the standard Hugging Face sections. It runs in your browser, so nothing you type is uploaded.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 1, 2026

Build your model cardREADME.md + YAML

Spec verified

Start from a preset

Model name*

Author / org

Short description (summary)71/300

License

Task (pipeline_tag)

Library

Base model

Language(s) — ISO 639-1

Training datasets

induwara/sinhala-news

Tags3/20

sinhalanewsdistilbert

Evaluation results

"Get started" snippet

Intended / direct use

Out-of-scope / misuse

Bias, risks & limitations

Training data

Citation (BibTeX)

Metadata checklist5/5

Model name set— The H1 and the model-index name come from this.
License chosen— The Hub filters and gates downloads by license.
Task (pipeline_tag) set— Drives discovery filters and the inference widget.
Summary written— The one-line lead under the H1; the first thing readers see.
At least one evaluation result— Adds a model-index block, unlocking the Hub's metrics badges.

README.mdValid frontmatter

---
license: apache-2.0
language:
- si
library_name: transformers
pipeline_tag: text-classification
base_model: distilbert-base-uncased
datasets:
- induwara/sinhala-news
tags:
- sinhala
- news
- distilbert
metrics:
- accuracy
model-index:
- name: sinhala-news-classifier
  results:
  - task:
      type: text-classification
    dataset:
      name: sinhala-news test
      type: sinhala-news-test
    metrics:
    - type: accuracy
      value: 0.912
---

# sinhala-news-classifier

DistilBERT fine-tuned to classify Sinhala news headlines into 6 topics.

## Model Details

### Model Description

- **Developed by:** induwara
- **Model type:** text-classification
- **Language(s):** si
- **License:** apache-2.0
- **Finetuned from model:** distilbert-base-uncased

## Uses

### Direct Use

Classify a Sinhala news headline into one of six topic labels.

### Downstream Use

[More Information Needed]

### Out-of-Scope Use

Not validated for long-form articles or languages other than Sinhala.

## Bias, Risks, and Limitations

Training data is drawn from a small set of publishers, so topic balance and regional coverage may be skewed.

### Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. [More Information Needed]

## How to Get Started with the Model

Use the code below to get started with the model.

```python
from transformers import pipeline

pipe = pipeline("text-classification", model="induwara/sinhala-news-classifier")
```

## Training Details

### Training Data

~40k labelled Sinhala headlines from the induwara/sinhala-news dataset.

### Training Procedure

[More Information Needed]

## Evaluation

### Testing Data, Factors & Metrics

[More Information Needed]

### Results

| Metric | Value | Dataset |
| --- | --- | --- |
| accuracy | 0.912 | sinhala-news test |

## Environmental Impact

Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in Lacoste et al. (2019).

- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]

## Technical Specifications

[More Information Needed]

## Citation

**BibTeX:**

[More Information Needed]

Generated entirely in your browser — nothing you type is uploaded.

How it works

On the Hugging Face Hub, a model repository's README.md is the model card. It has two parts: a YAML metadata block between two ---fences at the very top, and Markdown documentation below it. Get the YAML keys wrong and the Hub quietly ignores them, so the model page shows "no model card" and ranks nowhere. This tool is a deterministic assembler — it maps your form fields to the exact keys and the exact section order the Hub documents.

YAML frontmatter. Only the fields you fill in are emitted, in the Hub-recognised order: license, language, library_name, pipeline_tag, base_model, datasets, tags, metrics, and model-index. Empty keys are dropped so the block stays clean and valid. Values containing YAML-hostile characters (a colon, a hash) are quoted automatically so the block never breaks.

Evaluation → model-index. Each evaluation row you add (metric, value, dataset) becomes one entry under model-index → results, with task.type taken from your pipeline tag and a slugified dataset.type. Numeric values stay bare; non-numeric ones are quoted. With zero rows, both metrics and model-index are omitted — the correct behaviour, not an oversight.

Body. The generator emits # Model name, your summary as the lead paragraph, then the fixed H2 skeleton in Hugging Face's annotated-template order: Model Details, Uses (Direct / Downstream / Out-of-Scope), Bias Risks and Limitations, How to Get Started (a framework-specific code snippet), Training Details, Evaluation, Environmental Impact, Technical Specifications, and Citation. Sections you leave blank carry the [More Information Needed] placeholder — exactly what the official template does — so the card is complete and honest about its gaps.

Cross-check.The "spec verified" badge is not decorative. After each keystroke an independent function re-scans the emitted frontmatter and confirms every top-level key is in the recognised set, that key order matches the documented order, and that model-index appears if and only if there is at least one evaluation row. The keys, license list and section order come straight from the Hugging Face docs cited below, last verified 2026-07-01.

Worked examples

Text-classification model (Sinhala news)

Name: sinhala-news-classifier · Author: induwara
License: apache-2.0 · Language: si
Task: text-classification · Base: distilbert-base-uncased
Datasets: induwara/sinhala-news · Tags: sinhala, news, distilbert
Eval: accuracy = 0.912 on sinhala-news test

---
license: apache-2.0
language:
- si
library_name: transformers
pipeline_tag: text-classification
base_model: distilbert-base-uncased
datasets:
- induwara/sinhala-news
tags:
- sinhala
- news
- distilbert
metrics:
- accuracy
model-index:
- name: sinhala-news-classifier
  results:
  - task:
      type: text-classification
    dataset:
      name: sinhala-news test
      type: sinhala-news-test
    metrics:
    - type: accuracy
      value: 0.912
---

# sinhala-news-classifier

DistilBERT fine-tuned to classify Sinhala news headlines into 6 topics.
...

Every top-level key is Hub-recognised. The one eval row produces one model-index result; 0.912 stays a bare number because it parses as numeric; dataset.type is the slug of the dataset name (sinhala-news-test).

Text-generation LLM, minimal metadata

Name: tiny-lk-llm (no author)
License: mit · Task: text-generation
No datasets, no evaluation rows
Out-of-scope: Not for medical, legal, or financial advice.

---
license: mit
library_name: transformers
pipeline_tag: text-generation
tags:
- chat
- sri-lanka
---

# tiny-lk-llm

A 1.1B parameter chat model for Sri Lankan English.
...
### Out-of-Scope Use

Not for medical, legal, or financial advice.
...

Zero eval rows ⇒ metrics and model-index are correctly omitted. The filled Out-of-Scope section appears; every other blank section carries the [More Information Needed] placeholder.

Edge case — empty form

License: (cleared) · Task: (cleared)
Every field left blank

---
---

# my-awesome-model

[More Information Needed]
...

With nothing supplied, the frontmatter is the minimal valid ---\\n---. The card is still emitted so you can see the full skeleton; the checklist flags license, task and summary as missing.

Frequently asked questions

Sources & references

The recognised YAML keys, the license list and the section order on this page were last cross-checked against the Hugging Face docs on 2026-07-01. The format is a documented specification; this page is reviewed whenever Hugging Face revises the model card template.

Related tools

LiveAI

System Prompt Generator

Turn a short form into a structured system prompt — role, task, tone, output format, guardrails and few-shot examples — for ChatGPT Custom Instructions, a Custom GPT, or the Claude, OpenAI or Gemini system field. Generic Markdown or Claude XML style, with a live preview and copy/download. Runs entirely in your browser; nothing is uploaded, no AI rewriting.

Open tool

LiveAI

Perplexity Calculator

Compute language-model perplexity from token probabilities, cross-entropy loss, or log-likelihood, with nats and bits-per-token conversions. Step-by-step, matches PyTorch, runs entirely in the browser.

Open tool

LiveAI

Prompt Formatter

Paste rough notes and get a well-structured AI prompt with role, task, constraints, and output format. Works with GPT, Claude, Gemini. Deterministic templating — runs entirely in your browser.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want another card field supported?

Email me at [email protected] — most fixes ship within 24 hours.