Model Card

The complementary documentation standard for training data rather than models. Originating from the same research group (Gebru et al., 2018), datasheets detail a dataset's motivation, composition, collection process, preprocessing steps, and recommended uses. While a model card answers 'What does this model do?', a datasheet answers 'What was this model trained on?' The two documents together provide end-to-end transparency for an AI system's full provenance chain.

A formal, machine-readable inventory extending the Software Bill of Materials (SBOM) concept to AI systems. An AIBOM catalogs every component in the supply chain: datasets, pre-trained models, fine-tuning checkpoints, training pipelines, and evaluation harnesses. Model cards serve as the human-readable summary of what an AIBOM encodes programmatically, enabling automated compliance scanning and vulnerability tracking across enterprise AI portfolios.

The field of techniques that make opaque model decisions interpretable to humans. While a model card documents intended use and aggregate evaluation results, explainability methods like SHAP, LIME, and integrated gradients provide instance-level explanations for individual predictions. Model cards often reference which explainability techniques are compatible with the model, bridging the gap between static documentation and dynamic interrogation.

A family of W3C specifications defining a standardized data model for provenance interchange on the web. PROV represents provenance as a directed graph of entities, activities, and agents with well-defined relationships like wasGeneratedBy and wasDerivedFrom. Model cards can be serialized using PROV-O (the ontology) to make their claims machine-readable and interoperable across governance platforms, enabling automated provenance queries.

The alignment between a model's predicted probability of correctness and its actual empirical accuracy. A well-calibrated model that outputs 70% confidence should be correct exactly 70% of the time. Model cards typically report Expected Calibration Error (ECE) and reliability diagrams as key evaluation metrics. Poor calibration undermines trust even when raw accuracy is high, making this a critical transparency signal for high-stakes deployment decisions.

The systematic evaluation of a model's propensity to generate factually incorrect or nonsensical outputs presented with high confidence. Modern model cards increasingly include hallucination rates measured against grounded knowledge bases using metrics like factual consistency (FactCC) and attribution precision. For retrieval-augmented systems, hallucination risk is often stratified by whether the model had access to relevant retrieved context during generation.