Megatron-BERT-large Swedish 165k

This BERT model was trained using the Megatron-LM library. The size of the model is a regular BERT-large with 340M parameters. The model was trained on about 70GB of data, consisting mostly of OSCAR and Swedish newspaper text curated by the National Library of Sweden.

Training was done for 165k training steps using a batch size of 8k; the number of training steps is set to 500k, meaning that this version is a checkpoint. The hyperparameters for training followed the setting for RoBERTa.

The model has three sister models trained on the same dataset:

• ? BERT Swedish
• Megatron-BERT-base-600k
• Megatron-BERT-base-125k

and an earlier checkpoint

• Megatron-BERT-large-110k

Acknowledgements

We gratefully acknowledge the HPC RIVR consortium ( https://www.hpc-rivr.si ) and EuroHPC JU ( https://eurohpc-ju.europa.eu ) for funding this research by providing computing resources of the HPC system Vega at the Institute of Information Science ( https://www.izum.si ).