Model reporting and information sharing
Anthropic has been a longstanding advocate for model cards, which are used to share information about a model’s intended purpose and limitations, training data, and performance. Model cards can also include elements like algorithmic audits to evaluate a system’s components and data sheets. Such disclosures increase transparency and allow oversight into a model’s development and suitability for a particular use. Anthropic will be releasing model cards as part of all future model releases. Our most recent model card, the Claude 2 model card linked here, gives a direct sense of Anthropic’s views on the importance and contribution of model cards and similar transparency mechanisms.