License YOLO
Back to Ultralytics Glossary

Model Card

A model card documents a trained model's intended use, training data, evaluation results, and limitations, a format Google researchers proposed in 2018.

A model card is a short document published with a trained machine learning model that records what the model is for, how it was trained and evaluated, how it performs for different groups and conditions, and where it fails. Margaret Mitchell, Timnit Gebru, and colleagues at Google proposed the format in 2018 in the paper Model Cards for Model Reporting, presented at the FAT* 2019 conference. Model cards are now standard practice for transparency in AI: a model repository on Hugging Face shows its model card as its front page, and AI labs publish cards for major releases, such as the Claude 3 model card Anthropic released in 2024.

What a Model Card Contains#

The original paper proposed nine sections:

  • Model details: Developer, version, date, architecture, license, and where to send questions.
  • Intended use: The primary uses and users the model was built for, and uses that are out of scope.
  • Factors: The groups, instruments, and environments that could change performance, such as age group, camera type, or lighting.
  • Metrics: Which performance metrics are reported, at what decision thresholds, and why they were chosen.
  • Evaluation data: The datasets used for testing and why they were chosen.
  • Training data: The data the model learned from, described in as much detail as can be shared.
  • Quantitative analyses: Results broken down by each factor, alone and in combination.
  • Ethical considerations: Sensitive data, risks to people, and the mitigations applied.
  • Caveats and recommendations: Known gaps, such as conditions that were never tested.

Quantitative analyses are the core of the idea. A single headline score hides failures concentrated in one subgroup, so the paper's example card for a smiling detector trained on the CelebA face dataset reports error rates separately by age and gender, the disaggregated evaluation that exposes algorithmic bias and dataset bias. For an object detection model, the same section reports mean average precision (mAP) per class and per condition, such as small objects, night images, or each camera, along with inference latency on the target hardware.

Model Cards on Hugging Face#

On the Hugging Face Hub, a model card is the repository's README.md: Markdown sections for people, under a YAML metadata header that the Hub parses to filter models by task, license, and dataset and to display evaluation results. The huggingface_hub library builds both parts from structured data:

from huggingface_hub import EvalResult, ModelCard, ModelCardData
from torchvision.models import ResNet50_Weights

# Torchvision ships the published ImageNet accuracy with each set of weights
weights = ResNet50_Weights.IMAGENET1K_V2
top1 = weights.meta["_metrics"]["ImageNet-1K"]["acc@1"]

# Machine-readable metadata: library, training data, and evaluation results
result = EvalResult(
    task_type="image-classification",
    dataset_type="imagenet-1k",
    dataset_name="ImageNet-1K",
    metric_type="accuracy",
    metric_value=top1,
)
card_data = ModelCardData(
    model_name="resnet50", library_name="torchvision", datasets=["imagenet-1k"], eval_results=[result]
)

# Fill the default template: a YAML header followed by Markdown sections
card = ModelCard.from_template(card_data, model_summary="ResNet-50 image classifier trained on ImageNet-1K.")
print(card.data.to_yaml())
print([line[3:] for line in card.content.splitlines() if line.startswith("## ")])

The YAML output lists the training dataset and a model-index block recording 80.858% top-1 accuracy on ImageNet-1K, which the Hub displays as an evaluation result. The second line prints the template's sections, from Model Details and Uses through Bias, Risks, and Limitations to Evaluation and Environmental Impact, a direct descendant of the nine sections in the original paper.

  • Datasheets for datasets: Proposed by Timnit Gebru and colleagues in 2018, they document how a dataset was collected, labeled, and meant to be used. A model card points to the datasheet of its training data instead of repeating it.
  • System cards: Describe a deployed AI system made of several models, prompts, safety filters, and human processes. Meta introduced the term in 2022, and labs building large language models (LLMs), including OpenAI, now publish a system card with each major model release.
  • Model registry: Stores model files, versions, and lineage, while the card explains what each registered version is fit for.
  • Explainable AI (XAI): Explains why a model made an individual prediction, while a card describes behavior across a whole evaluation set.

Regulation has made the format a requirement in some settings. Under the EU AI Act, providers of general-purpose AI models must, since August 2, 2025, keep technical documentation of training and evaluation and give the developers who build on their models information on capabilities and limitations. The Commission's General-Purpose AI Code of Practice supplies a Model Documentation Form for it, with card-like fields such as intended uses, training data, and energy consumption. Cards are part of the wider practice of AI governance and responsible AI.

Keeping a Card Current#

A card describes one model version, so it changes whenever the model, its data, or its deployment conditions change. Production model monitoring supplies the evidence: data drift or a new failure mode becomes a caveat in the next version's card.

Part of a card can also travel inside the model file. When Ultralytics YOLO26 models are exported to formats such as ONNX or OpenVINO, the exporter embeds metadata with a description naming the training dataset, the author, export date, package version, license, task, input size, and class names, so a deployed copy still identifies what it is and what it detects.

Explore solutions

Computer vision in sports

Computer vision in sports

Discover how Ultralytics YOLO powers sports vision with player and ball tracking, performance analytics, coaching insights, and athlete rehabilitation.
Learn more
Computer vision for aerial imagery

Computer vision for aerial imagery

Turn drone and aerial imagery into real-time insights for agriculture, aquaculture, conservation, and geospatial analysis with Ultralytics YOLO.
Learn more
Computer vision in aerospace

Computer vision in aerospace

Apply computer vision in aerospace with Ultralytics YOLO for aircraft inspection, runway FOD detection, airport ground operations, and composite defect detection.
Learn more
Computer vision in security

Computer vision in security

Protect every site with Ultralytics YOLO models. Vision AI powers perimeter monitoring, threat detection, license-plate recognition, and workplace safety.
Learn more
Computer vision in robotics

Computer vision in robotics

Power smarter machines with Ultralytics YOLO models. Vision AI in robotics drives autonomous navigation, perception, object tracking, and real-time control.
Learn more
Computer vision in logistics

Computer vision in logistics

Streamline logistics with Ultralytics YOLO models. Vision AI enables package inspection, sorting, vehicle tracking, and real-time warehouse safety monitoring.
Learn more
Computer vision in retail

Computer vision in retail

Reimagine retail with Ultralytics YOLO models. Vision AI powers inventory tracking, shelf monitoring, queue management, and smarter customer insights.
Learn more
Computer vision in healthcare

Computer vision in healthcare

Build healthcare solutions with Ultralytics YOLO models. Vision AI in healthcare powers faster medical imaging, smarter diagnostics, and patient monitoring.
Learn more
Computer vision in manufacturing

Computer vision in manufacturing

Optimize manufacturing with Ultralytics YOLO models. Vision AI drives quality control, defect detection, PPE compliance, and assembly line automation.
Learn more
Computer vision in automotive

Computer vision in automotive

Apply computer vision in automotive with Ultralytics YOLO models. Vision AI elevates road safety, driver assistance, and vehicle automation for smarter roads.
Learn more
Computer vision in agriculture

Computer vision in agriculture

Bring vision AI to smart agriculture with Ultralytics YOLO models. Power crop monitoring, livestock tracking, and precision farming for higher, smarter yields.
Learn more

Let's build the future of AI together!

Begin your journey with the future of machine learning