andreasmadsen/

efficient_mlm_m0.20

This is a model checkpoint for "Should You Mask 15% in Masked Language Modeling" (code).

The original checkpoint is avaliable at princeton-nlp/efficient_mlm_m0.20. Unfortunately this checkpoint depends on code that isn't part of the official transformers library. Additionally, the checkpoints contains unused weights due to a bug.

This checkpoint fixes the unused weights issue and uses the RobertaPreLayerNorm model from the transformers library.