DeBERTa trained from scratch
continued training from https://huggingface.co/mikesong724/deberta-wiki-2006
Source data: https://dumps.wikimedia.org/archive/2010/
Tools used: https://github.com/mikesong724/Point-in-Time-Language-Model
2010 wiki archive 6.1 GB trained 18 epochs = 108GB + 2006 (65GB)
GLUE benchmark
cola (3e): matthews corr: 0.3640
sst2 (3e): acc: 0.9106
mrpc (5e): F1: 0.8505, acc: 0.7794
stsb (3e): pearson: 0.8339, spearman: 0.8312
qqp (3e): acc: 0.8965, F1: 0.8604
mnli (3e): acc_mm: 0.8023
qnli (3e): acc: 0.8889
rte (3e): acc: 0.5271
wnli (5e): acc: 0.3380