---
language:
  - he
  - arc
license: mit
tags:
  - punctuation
  - talmud
  - hebrew
  - aramaic
  - token-classification
base_model: dicta-il/BEREL_3.0
pipeline_tag: token-classification
---

# Talmud Punctuator — Model B (Steinsaltz)

Fine-tuned [BEREL 3.0](https://huggingface.co/dicta-il/BEREL_3.0) for predicting punctuation
in Talmudic Aramaic/Hebrew text.

**Model B** reflects the Steinsaltz/William Davidson Edition punctuation style,
trained on a 15,000-sentence sample across 36 masekhtot of the Babylonian Talmud.

This is a preliminary model. For better accuracy, retrain on the full 80K-sentence
dataset using the included Colab notebook (`train_on_colab.ipynb`).

## Training details

- **Base model**: BEREL 3.0 (`dicta-il/BEREL_3.0`)
- **Head**: Linear classification (768 → 8 labels)
- **Data**: 15,000 sentences sampled from 36 masekhtot (~80K available)
- **Epochs**: 3, Batch size: 16, LR: 2e-5
- **Final loss**: 0.2148

## Labels

| Label | Meaning |
|-------|---------|
| `O` | No punctuation |
| `,` | Comma |
| `.` | Period |
| `:` | Colon |
| `;` | Semicolon |
| `?` | Question mark |
| `!` | Exclamation mark |
| `—` | Em-dash |

## Usage

Use with the `punctuator.py` script from [mivami](https://github.com/joshuawaxman/mivami).

## Upgrading to full dataset

Use the Colab notebook to train on all 80K sentences with a free GPU:
1. Set `CONTINUE_FROM = "Joshua2/talmud-punctuator-B"` to start from this checkpoint
2. Upload `Steinsaltz_combined.txt`
3. Run on a T4 GPU (~3-5 hours)