Evidence item

Under data constraints, optimal scaling inverts Chinchilla's equal-scaling advice. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 7351

Source: Pretraining data science (blog) · locator {} · excerpt hash sha256:06ae9933e7fb74188843719249daa89c70b457cb79b16033947350e0fa0cc0a4

Exact excerpt

Because excess parameters lose value faster than repeated data, under a data constraint you should spend surplus compute on epochs before parameters, which directly inverts Chinchilla's equal-scaling advice.

Linked claims

Claim	Direction	Relevance	Steps	Counted
a48d8ddf	SUPPORT	STRONG	1	yes

Audits

None.