This model is a fine-tuned version of BERiT on the COHeN dataset. It achieves the following results on the evaluation set:
Loss: 0.4418
Accuracy: 0.8622
Model Description
COHeN (Classification of Old Hebrew via Neural Net) is a text classification model for Biblical Hebrew that assigns Hebrew texts to one of four chronological phases: Archaic Biblical Hebrew (ABH), Classical Biblical Hebrew (CBH), Transitional Biblical Hebrew (TBH), or Late Biblical Hebrew (LBH). It allows scholars to check their intuition regarding the dating of particular verses.
How to Use
from transformers import AutoModelForSequenceClassification, AutoTokenizer
model_name = 'gngpostalsrvc/COHeN'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
Training Procedure
COHeN was trained on the COHeN dataset for 20 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: 0.0027
weight_decay: 0.0049
train_batch_size: 32
eval_batch_size: 32
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
