Training harness to pretrain a Marian model using mBART
Provides the mBART pretraining strategy for Marian neural machine translation models.
Implemented using an external training harness that reads monolingual data, applies mBART noise and sends it to a Marian training process using a pair of named pipes.
Used for the University of Edinburgh WMT2020 English-Tamil News Translation task submission and for the GoURMET project English-Tamil machine translation models.