FRAME orders pretraining data Q3->Q4->Q1->Q2 by PPL and PPL-difference quadrants and reports 15.3/18.2 percentage-point MMLU/CMMLU gains over random ordering for a 3B model.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
FRAME: Boosting LLMs with A Four-Quadrant Multi-Stage Pretraining Strategy
FRAME orders pretraining data Q3->Q4->Q1->Q2 by PPL and PPL-difference quadrants and reports 15.3/18.2 percentage-point MMLU/CMMLU gains over random ordering for a 3B model.