Pretrained model on Dhivehi language using masked language modeling (MLM).
Tokenizer
The WordPiece tokenizer uses several components:
Normalization: lowercase and then NFKD unicode normalization.
Pretokenization: splits by whitespace and punctuation.
Postprocessing: single sentences are output in format [CLS] sentence A [SEP] and pair sentences in format [CLS] sentence A [SEP] sentence B [SEP].
Training
Training was performed over 16M+ Dhivehi sentences/paragraphs put together by @ashraq. An Adam optimizer with weighted decay was used with following parameters:
