# dv-wave

This is my third version of a Dhivehi language model trained with
Google Research's [ELECTRA](https://github.com/google-research/electra).

Tokenization and pre-training CoLab: https://colab.research.google.com/drive/1ZJ3tU9MwyWj6UtQ-8G7QJKTn-hG1uQ9v?usp=sharing

Using SimpleTransformers to classify news https://colab.research.google.com/drive/1KnyQxRNWG_yVwms_x9MUAqFQVeMecTV7?usp=sharing

V1: similar performance to mBERT on news classification task after finetuning for 3 epochs (52%)

V2: fixed tokenizers do_lower_case=False and strip_accents=False to preserve vowel signs of Dhivehi

> 8-topic news classification score: 88.6% compared to mBERT: 51.8%

V3: trained longer on larger corpus (added OSCAR and Wikipedia)

> news classification score: 91.9%

## Corpus

Trained on @Sofwath's 307MB corpus of Dhivehi text: https://github.com/Sofwath/DhivehiDatasets

Plus [OSCAR](https://oscar-corpus.com/) and 2020-10-01 dump of dv.wikipedia.org

## Vocabulary

Included as vocab.txt in the upload