BERiT

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

• Loss: 3.9931

Model description

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.

How to Use

from transformers import RobertaModel, RobertaTokenizerFast

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

Training procedure

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

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 150

Framework versions

• Transformers 4.24.7
• Pytorch 2.0.0+cu118
• Datasets 2.11.0
• Tokenizers 0.13.3