TiRoBERTa: RoBERTa Pretrained for the Tigrinya Language

We pretrain a RoBERTa base model for Tigrinya on a dataset of 40 million tokens trained for 40 epochs.

Contained in this repo is the original pretrained Flax model that was trained on a TPU v3.8 and it's corresponding PyTorch version.

Hyperparameters

The hyperparameters corresponding to model sizes mentioned above are as follows:

Model Size	L	AH	HS	FFN	P	Seq
BASE	12	12	768	3072	125M	512

(L = number of layers; AH = number of attention heads; HS = hidden size; FFN = feedforward network dimension; P = number of parameters; Seq = maximum sequence length.)

Framework versions

Transformers 4.12.0.dev0
Pytorch 1.9.0+cu111
Datasets 1.13.3
Tokenizers 0.10.3

Citation

If you use this model in your product or research, please cite as follows:

@article{Fitsum2021TiPLMs,
  author={Fitsum Gaim and Wonsuk Yang and Jong C. Park},
  title={Monolingual Pre-trained Language Models for Tigrinya},
  year=2021,
  publisher={WiNLP 2021 at EMNLP 2021}
}