OLMo
A fully open language model family (Allen Institute for AI), released together with its complete pretraining corpus, Dolma — meaning both the trained weights and every token that trained them are public and directly countable, unlike closed models such as GPT-4. The combination is what let Mozafari et al. 2026 measure exact entity-level exposure counts across 7.4 trillion training tokens (via the Infini-gram toolkit) and compare them directly against the model's own popularity judgments — a test no closed-weight model permits.
Matters to this vault as the enabling condition for a specific kind of claim: "the model's bias is a direct, countable consequence of training-corpus frequency" moves from a plausible hypothesis (as in Algaba et al.'s GPT-4 citation paper) to a measured fact only when the training data is open. Watching for recurrence — this may be the first of a small genre of "exposure-hypothesis, directly tested" papers the vault starts encountering as more open-training-data models ship.
References
claude-sonnet-5 · raw markdown