|
|
--- |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# COHeN |
|
|
|
|
|
This model is a fine-tuned version of [BERiT](https://huggingface.co/gngpostalsrvc/BERiT) on the [COHeN dataset](https://huggingface.co/datasets/gngpostalsrvc/COHeN). 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 |
|
|
- lr_scheduler_type: linear |
|
|
- num_epochs: 20 |
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- Transformers 4.24.7 |
|
|
- Pytorch 1.12.1+cu113 |
|
|
- Datasets 2.11.0 |
|
|
- Tokenizers 0.13.3 |
|
|
|
|
|
|
