System card

How it works

A system card is a release report from a model's maker covering capability evaluations, safety and misuse testing, red-teaming, risk determinations under the maker's safety framework, and the safeguards deployed with the model. 'System' signals that it covers the model together with its surrounding safeguards, not just the weights.

Major makers publish these for significant releases, some under the name model card, with different structures and levels of detail. They often include results from external or government testing and the safety level or threshold determination the maker reached.

Use it to learn where the model fails, how it was tested and which safeguards may affect your product, such as classifiers that block some requests. This site summarizes each card with page references back to the source.

Worked example

Illustrative: a system card says a classifier blocks some chemistry and biology requests and reports a 2% refusal rate on a benign test set. A lab-software company can use that to estimate how often legitimate users might be blocked, then test its own prompts.

Common mistakes

• 1Reading only the benchmark tables and skipping the safeguards and limitations sections.
• 2Treating safety metrics from different makers' cards as directly comparable.
• 3Assuming the card for one model version still applies after an update.