Same pre-training loss, better downstream: Implicit bias matters for language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1