This model is a fine-tuned version of roberta-base on the Tanakh dataset. It achieves the following results on the evaluation set:

BERiT is a masked-language model for Biblical Hebrew, a low-resource ancient language preserved primarily in the text of the Hebrew Bible. Building on the work of Sennrich and Zhang (2019) and Wdowiak (2021) on low-resource machine translation, it employs a modified version of the encoder block from Wdowiak’s Seq2Seq model. Accordingly, BERiT is much smaller than models designed for modern languages like English. It features a single attention block with four attention heads, smaller embedding and feedforward dimensions (256 and 1024), a smaller max input length (128), and an aggressive dropout rate (.5) at both the attention and feedforward layers.

The BERiT tokenizer performs character level byte-pair encoding using a 2000 word base vocabulary, which has been enriched with common grammatical morphemes.

from transformers import RobertaModel, RobertaTokenizerFast

BERiT_tokenizer = RobertaTokenizerFast.from_pretrained('gngpostalsrvc/BERiT')
BERiT = RobertaModel.from_pretrained('gngpostalsrvc/BERiT')

BERiT was trained on the Tanakh dataset for 150 epochs using a Tesla T4 GPU. Further training did not yield significant improvements in performance.

The following hyperparameters were used during training: