Advancement or Acceleration? How Data Curation Improves LLM Training
It is shown that advancement requires a correctable train-target gap, and the condition is operationalized as the $\Delta$-test, an observable discrepancy between the uncurated baseline's training and target losses that is computable prior to expensive curation.
Jia-Xing Wang, Zi-Cheng Zhang, De-Ping Xiang et al.
· 0 citations