German cybersecurity agency releases quality criteria for AI training data

BSI is Germany's Federal Office for Information Security. They published a german draft document, an english version will follow shortly.

Translated blurb:

Their new draft - in German - of the QUAIDAL criteria catalogue offers a comprehensive evaluation framework for quality criteria for training data in the AI lifecycle. It aims to systematically record and ensure data quality, transparency, traceability, fairness and compliance.

Based on common norms and standards, the catalogue defines ten central quality criteria, which are then divided into four levels: abstract criteria, concrete data components, operational measures and associated metrics and methods. In total, the ten criteria are mapped by 143 metrics and methods, enabling a detailed and holistic assessment of data quality. The content covers the entire process - from data acquisition and cleansing to bias detection and mitigation, data protection, governance and audits. Relevant sources of information and source codes are also available.

The requirements are based on international standards and support the implementation of the AI Regulation. The catalogue is aimed at anyone who plans, develops or reviews training data for AI projects. As QUAIDAL is designed as a German guideline, the document is primarily available in German; an English version will follow shortly.