LVBERT is a Latvian BERT model, trained on 0.5 billions Latvian tokens from Latvian Balanced corpus, Wikipedia, comments and articles from various news portals.
LVBERT uses BERT-Base configuration with 12 trasformer layers, each of 768 hidden units, 12 heads, 128 sequence length, 128 mini-batch size. It was trained using original TensorFlow code for BERT with the whole-word masking and the next sentence prediction objectives.
Sentencepiece model tokenizes text into subword tokens that are used by the the BERT model. The file "vocab.txt" lists all the subword tokens. Tokens prepended with "##" are only used for continuation, ie. they're necessarily combined with a preceding token(s) to form a full word. The total vocabulary of the model is composed of 32000 subword tokens.
For practical use, LVBERT model is available at Hugging Face: https://huggingface.co/AiLab-IMCS-UL/lvbert
LVBERT model is published under GNU General Public Licence (version 3) license. . . .
