megatron.bert-large.unigram-64k-pretok.500k-steps

megatron.bert-large.unigram-64k-pretok.500k-steps

This BERT model was trained using the NeMo library. The size of the model is a regular bert-large. The model was trained on more than 245GB of data, consisting mostly of web-data and Swedish newspaper text curated by the National Library of Sweden.

Training was done for 500k training steps using a batch size of 8k.

The model has multiple sibling models trained on the same dataset using different tokenizers or more/less parameters:

• megatron.bert-base.bpe-32k-no_pretok.25k-steps
• megatron.bert-base.bpe-64k-no_pretok.25k-steps
• megatron.bert-base.spe-bpe-32k-no_pretok.25k-steps
• megatron.bert-base.spe-bpe-32k-pretok.25k-steps
• megatron.bert-base.spe-bpe-64k-no_pretok.25k-steps
• megatron.bert-base.spe-bpe-64k-pretok.25k-steps
• megatron.bert-base.unigram-32k-no_pretok.25k-steps
• megatron.bert-base.unigram-32k-pretok.25k-steps
• megatron.bert-base.unigram-64k-no_pretok.25k-steps
• megatron.bert-base.unigram-64k-pretok.25k-steps
• megatron.bert-base.wordpiece-32k-no_pretok.25k-steps
• megatron.bert-base.wordpiece-32k-pretok.25k-steps
• megatron.bert-base.wordpiece-64k-no_pretok.25k-steps
• megatron.bert-base.wordpiece-64k-pretok.25k-steps
• megatron.bert-large.bpe-64k-no_pretok.25k-steps
• megatron.bert-large.spe-bpe-32k-pretok.25k-steps
• megatron.bert-large.unigram-32k-pretok.25k-steps
• megatron.bert-large.unigram-64k-pretok.25k-steps
• megatron.bert-large.wordpiece-32k-pretok.25k-steps
• megatron.bert-large.wordpiece-64k-pretok.25k-steps

Acknowledgements

The training was performed on the Luxembourg national supercomputer MeluXina. The authors gratefully acknowledge the LuxProvide teams for their expert support.