Data Provenance

Knowing what is actually in a training corpus, and where each part came from. It sounds like bookkeeping and turns out to be research: web-scale corpora are assembled from crawls and dumps whose contents nobody has read, so what a model trained on is genuinely unknown to the people who trained it until somebody goes and looks.

When people do look, they find things that were not supposed to be there — personal data, benchmark test sets, machine-translated text presented as native, whole books. The finding that matters most for everything downstream is that these are not rare contaminants but structural features of web crawls, so a pipeline that assumes a clean corpus and filters for quality is filtering something it has not characterised.

Provenance is a property of a document that has to be recorded when the document is collected, because it cannot be recovered afterwards. Once a corpus is deduplicated, shuffled and tokenised, the mapping back to sources is gone, and reconstructing it costs more than keeping it would have. So the share of a corpus with known origin is fixed at collection time and only ever falls as the corpus is processed — which is why retrofitting provenance onto an existing dataset does not work.