AI Model Card & Dataset Card Generator (README with Metadata)
A guided form for model and dataset cards, with the metadata the Hub reads.
Try before you buy.
- Free preview: the rendered card with its first three sections under a watermark, how complete it is, and what the Hub would refuse in its metadata.
- Locked until you unlock it: download, copy and print.
- Unlock: Pro pass, ₹179 for 30 days or ₹79 for 1 week, a one-time payment that never renews. Unlock: Pro pass, $5.99 for 30 days or $2.99 for 1 week, a one-time payment that never renews.
Ways to unlock shows how to get the full result.
Printing this result is locked in the free preview.
Locked in the free preview. Opens the ways to unlock this result.
Locked in the free preview. Batch runs unlock with a pass.
Locked in the free preview. Query results unlock with a pass.
About the AI Model Card & Dataset Card Generator (README with Metadata)
Write the card that goes with a machine-learning model or dataset: the README.md the Hugging Face Hub shows on its page. Fill in a guided form and the card writes itself as you type, in the section layout of the Hub’s own model card and dataset card templates, with the YAML metadata at the top that the Hub reads for search and filters: the language, licence, base model, task, tags, datasets and, for models, an evaluation table that also becomes the model-index block of results.
The form also estimates the energy and emissions of training (devices × power × hours × PUE × the grid’s carbon intensity, the method behind the ML CO2 Impact calculator) and writes them as the co2_eq_emissions metadata, shows how complete the card is, and lists what the Hub would refuse. It can fill fields from your own files: a Transformers config.json, the trainer_state.json the Trainer saves, a CSV of results, or a model file’s header. The free preview shows the rendered card with its first three sections under a watermark; every section, the metadata block, the imports and the README.md download, copy and print come with a Pro pass or an unlock.
How to use it
- Choose Model card or Dataset card. Press Fill in an example to see a finished card, or start with your own.
- Fill in Basics: the Hub ID (owner/name), a summary, who made it, the languages and the licence. For a model add its type, base model, library and task; for a dataset its task categories and size.
- Work through the other sections. Empty recommended parts read [More Information Needed] in the card, and the meter above the card lists them.
- For a model, add each evaluation result as a row (task, dataset ID and name, split, metric and value): the rows become the results table and the model-index metadata.
- Under Environmental impact, choose the hardware (its maximum power fills in), the number of devices, the hours, the PUE and the grid’s carbon intensity to estimate the emissions.
- Check the card on the right. With a Pro pass or an unlock, read the Markdown and the metadata, import your files, then download README.md, copy it or print it as a PDF.
- Press Save the form to keep a JSON copy of what you typed, and Open a saved form to go on later.
Examples
Results: text-classification · stanfordnlp/imdb (IMDB) · plain_text · test · accuracy 0.9312 (Test accuracy) text-classification · stanfordnlp/imdb (IMDB) · plain_text · test · f1 0.9301 text-classification · sst2 (SST-2) · validation · accuracy 91.0 ± 0.4
model-index:
- name: example-org/tiny-sentiment
results:
- task: { type: text-classification, name: Text Classification }
dataset: { name: IMDB, type: stanfordnlp/imdb, config: plain_text, split: test }
metrics: [ accuracy 0.9312 (Test accuracy), f1 0.9301 ]
- task: { type: text-classification, name: Text Classification }
dataset: { name: SST-2, type: sst2, split: validation }
metrics: [ accuracy "91.0 ± 0.4" ]Rows with the same task, dataset, config and split share one entry, exactly as the huggingface_hub library groups them.
Hardware: 8 × NVIDIA A100 SXM4 80 GB (400 W each) Hours: 2.5 · PUE: 1.1 · Carbon intensity: 0.432 kg CO2eq/kWh
Energy: 8 × 400 W × 2.5 h × 1.1 = 8.8 kWh Emissions: 8.8 kWh × 0.432 = 3.802 kg CO2eq Metadata: co2_eq_emissions.emissions = 3801.6 (grams, as the Hub expects)
Common uses
- Publishing a fine-tuned model on the Hugging Face Hub with a complete card instead of the default stub.
- Documenting a dataset’s collection, annotation, personal data and limitations before sharing it.
- Reporting evaluation results in a structured form that leaderboards and search can read.
- Estimating and declaring the carbon emissions of a training run.
- Preparing the documentation an organisation or a regulator asks for before a model is used.
What the metadata block holds
language: ISO 639-1, 639-2 or 639-3 codes, orcodeandmultilingual.license: an identifier from the Hub’s list (such asapache-2.0,mit,cc-by-4.0); for any other licence,otherwithlicense_nameandlicense_link.- Models:
library_name,tags,datasets,metrics,base_model,pipeline_tag,co2_eq_emissions(in grams) andmodel-index(task type, dataset type and name, metric type and value, with optional config, split and names). - Datasets:
pretty_name,tags,task_categories,size_categories,annotations_creators,language_creators,multilingualityandsource_datasets.
How the emissions are estimated
Energy (kWh) = the number of devices × the power of each (W) × the hours of training × the PUE (the data centre’s power usage effectiveness: 1 when the overhead is unknown) ÷ 1000. Emissions (kg CO2eq) = energy × the carbon intensity of the electricity grid (kg CO2eq per kWh). This is the method Lacoste et al. describe and the ML CO2 Impact calculator uses, which leaves the PUE for you to multiply in. The listed power of each part is its maximum (TDP), so the result is an upper estimate; a measured average power gives a closer figure.
What the imports read
- config.json (Transformers): the architecture, layer count, hidden size, attention and key-value heads, vocabulary, context length, weight precision and library version.
- trainer_state.json (the Trainer’s checkpoint state): epochs, steps, batch size, peak learning rate, training runtime (it fills the hours used), the final training loss and the last evaluation’s metrics.
- Results CSV: one row per result, with metric and value columns at least.
- A model file: the header of a GGUF, safetensors, ONNX or PyTorch file, read with the parsers of AI Model File Viewer: architecture, parameters, storage type and context length.
An import fills only fields that are still empty and merges tags; nothing you typed is overwritten.
Sources
- Hugging Face Hub: model cards, dataset cards and carbon emissions metadata
- The metadata specifications modelcard.md and datasetcard.md, and the list of licences
- The card templates of huggingface_hub and the task list of huggingface.js
- Lacoste et al., Quantifying the Carbon Emissions of Machine Learning, and the ML CO2 Impact calculator with its hardware table
- Mitchell et al., Model Cards for Model Reporting; Gebru et al., Datasheets for Datasets
Limitations
- The card follows the Hugging Face Hub’s format. Other model registries may name their fields differently.
- The Hub checks the metadata again when you push the README; the checks here cover the licence, languages, IDs, tasks, results and numbers, not every field the Hub accepts.
- The emissions estimate uses the hardware’s maximum power and the carbon intensity you enter; it does not measure anything.
- The text you write is used as it is: nothing is rewritten or invented.
Privacy
On this device: nothing is sent
Everything happens in your browser. What you enter or open here is not uploaded or stored by MySmartCoPilot.
Frequently asked questions
What do I get without a pass?
Without a pass, AI Model Card & Dataset Card Generator (README with Metadata) shows the rendered card with its first three sections under a watermark, how complete it is, and what the Hub would refuse in its metadata. Until you unlock it, the result can’t be downloaded, copied or printed. A Pro, Premium or Ultimate pass, a one-time payment that never renews, unlocks the full result. The pricing page lists the passes and their prices.
Where do I put the card?
Save it as README.md in the root of the model or dataset repository. The Hub reads the YAML block between the --- lines at the top and shows the rest as the card.
Why are emissions in grams?
The Hub’s co2_eq_emissions metadata takes grams of CO2 equivalent. The card’s text shows kilograms and the energy in kWh; the metadata holds the same figure in grams.
Does the page change what I write?
No. Your text goes into the card as you typed it; empty recommended parts read [More Information Needed], as in the Hub’s own templates, and empty optional parts are left out.
What does the free preview show?
The rendered card with its first three sections under a watermark, how complete it is and the metadata checks. The other sections, the Markdown and metadata views, the emissions estimate, the imports, and downloading, copying or printing the card come with a Pro pass or an unlock. Saving and opening the form as JSON always works.
Is anything uploaded?
No. The card is written in your browser, and imported files, including model files, are read on your device.