system card
Documentation that describes a model's capabilities, limitations, testing methodology, and known failure modes.
Think of it like
A nutritional label on a food product—ingredients, allergens, serving size.
Example
A system card might say: 'Claude 3 Opus: tested on MMLU with 88.7% accuracy, struggles with obscure foreign languages, refusal rate 2%.'
How it actually works
System cards come from responsible AI practices. They're meant to inform users about what to expect and what to watch for. No standard format; varies by lab. Good ones include: capability claims with citations, limitations (what doesn't work), evaluation results, known harm categories.
For product teams
Reduces risk of misuse; sets user expectations; helps compliance teams assess safety.
For engineers
Transparency tool; should include eval results and limitations.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome