Training data transparency discloses what a dataset contains, where it came from, who it represents, its limitations, and how it should be used. It helps researchers and auditors examine how bias may enter AI systems, although complete disclosure remains uncommon. section 3.2
Useful transparency practices highlight:
- Dataset contents and metadata section 3.2
- Demographic representation and possible limitations section 3.2
- Provenance and considerations for downstream use section 3.2
- Documentation and examination by researchers, auditors, and regulators section 3.2
From search tools to nutritional labels
A search tool for ROOTS, a 1.6-terabyte multilingual corpus used to train BLOOM, supports examination of training data. The Data Nutrition Project adapts food-label ideas to identify dataset ingredients such as metadata and demographic representation. section 3.2
