Model card
Overview
A model card is a short, standardized document accompanying a published machine learning model that provides essential factual information about the model: what it was designed for, what data it was trained on, how it performs across demographic groups and use cases, and what its known limitations and risks are. The concept was proposed by Mitchell et al. (2019) as a mechanism for model transparency and accountability.[1]
Model cards are produced by model developers (AI labs, research groups, enterprises) and published alongside the model release, typically on model hubs such as Hugging Face or in technical documentation. They are voluntary in most contexts but increasingly referenced in AI governance frameworks.
Typical sections
| Section | Content |
|---|---|
| Model details | Name, version, type, release date, developer |
| Intended use | Primary use cases, out-of-scope uses |
| Training data | Data sources, collection method, known composition issues |
| Evaluation results | Performance on benchmarks broken out by relevant subgroups |
| Limitations | Known failure modes, biases, technical constraints |
| Ethical considerations | Risk categories, mitigation measures |
| Caveats and recommendations | Context-specific guidance for deployment |
There is no enforced standard schema; different organizations use different formats. Hugging Face's model card template, the original Mitchell et al. structure, and newer EU AI Act–oriented disclosure forms are common variants.
| Document | Focus |
|---|---|
| Model card | Model transparency: capabilities, limitations, intended use |
| Datasheet for datasets | Dataset provenance, collection method, composition (analogous concept for data) |
| System card | System-level transparency (e.g., Anthropic's Claude system cards): multi-model pipelines and deployment-level risks |
| Technical report | Detailed pre-training and evaluation methodology; more extensive, research-audience |
Regulatory context
The EU AI Act (2024) imposes disclosure obligations on high-risk AI systems that overlap substantially with model card content. Model cards are increasingly referenced as a baseline disclosure artifact even where not yet legally mandated.
See also
References
- ↑ Mitchell, Margaret et al. "Model Cards for Model Reporting." FAccT 2019. https://arxiv.org/abs/1810.03993
