Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior

Benchmark Model Rank Results
question-answering-on-copa-zero-shotPPT (k-Shuffle Dyck), 3B, Marin, 100B PT tokens#3Accuracy: 85.0
language-modelling-on-lambadaPPT (k-Shuffle Dyck), 3B, Marin, 100B PT tokens#16Accuracy: 74.3
reading-comprehension-on-racePPT (k-Shuffle Dyck), 3B, C4, 100B PT tokens#30Accuracy: 37.2
question-answering-on-piqaPPT (k-Shuffle Dyck), 3B, C4, 100B PT tokens#87Accuracy: 77.1
sentence-completion-on-hellaswagPPT (k-Shuffle Dyck), 3B, C4, 100B PT tokens#128Accuracy: 66.3