language: dv | |
# dv-wave | |
This is a second attempt at 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 | |
dv-wave: 89% to mBERT: 52% | |
## Corpus | |
Trained on @Sofwath's 307MB corpus of Dhivehi text: https://github.com/Sofwath/DhivehiDatasets - this repo also contains the news classification task CSV | |
[OSCAR](https://oscar-corpus.com/) was considered but has not been added to pretraining; as of | |
this writing their web crawl has 126MB of Dhivehi text (79MB deduped). | |
## Vocabulary | |
Included as vocab.txt in the upload - vocab_size is 29874 | |