This model is a fine-tuned MarianMT model for Aramaic text diacritization (vocalization), converting consonantal Aramaic text to fully vocalized text with nikkud (vowel points).

Model Description

Model type: MarianMT (Encoder-Decoder Transformer)
Language: Aramaic (arc2arc)
Task: Text diacritization/vocalization
Base model: Helsinki-NLP/opus-mt-afa-afa
Parameters: 61,924,352 (61.9M)

Model Architecture

Architecture: MarianMT (Marian Machine Translation)
Encoder layers: 6
Decoder layers: 6
Hidden size: 512
Attention heads: 8
Feed-forward dimension: 2048
Vocabulary size: 33,714
Max sequence length: 512 tokens
Activation function: Swish
Position embeddings: Static

Training Details

Training Configuration

Training data: 12,110 examples
Validation data: 1,514 examples
Batch size: 8
Gradient accumulation steps: 2
Effective batch size: 16
Learning rate: 1e-5
Warmup steps: 1,000
Max epochs: 100
Training completed at: Epoch 36.33
Mixed precision: FP16 enabled

Training Metrics

Final training loss: 0.283
Training runtime: 21,727 seconds (~6 hours)
Training samples per second: 55.7
Training steps per second: 3.48

Evaluation Results

Test Set Performance

BLEU Score: 72.90
Character Accuracy: 63.78%
Evaluation Loss: 0.088
Evaluation Runtime: 311.5 seconds
Evaluation samples per second: 4.86

Usage

Basic Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
model_name = "johnlockejrr/aramaic-diacritization-model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Example input (consonantal Aramaic text)
consonantal_text = "בקדמין ברא יי ית שמיא וית ארעא"

# Tokenize input
inputs = tokenizer(consonantal_text, return_tensors="pt", max_length=512, truncation=True)

# Generate vocalized text
outputs = model.generate(**inputs, max_length=512, num_beams=4, early_stopping=True)

# Decode output
vocalized_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(f"Input: {consonantal_text}")
print(f"Output: {vocalized_text}")

Using the Pipeline

python

from transformers import pipeline

diacritizer = pipeline("text2text-generation", model="johnlockejrr/aramaic-diacritization-model")

# Process text
consonantal_text = "בראשית ברא אלהים את השמים ואת הארץ"
vocalized_text = diacritizer(consonantal_text)[0]['generated_text']
print(vocalized_text)

Training Data

The model was trained on a custom Aramaic diacritization dataset with the following characteristics:

Source: Consonantal Aramaic text (without vowel points)
Target: Vocalized Aramaic text (with nikkud/vowel points)
Data format: CSV with columns: consonantal, vocalized, book, chapter, verse
Data split: 80% train, 10% validation, 10% test
Text cleaning: Preserves nikkud in target text, removes punctuation from source

Data Preprocessing

Input cleaning: Removes punctuation and formatting while preserving letters
Target preservation: Maintains all nikkud (vowel points) and diacritical marks
Length filtering: Removes sequences shorter than 2 characters or longer than 1000 characters
Duplicate handling: Removes exact duplicates to prevent data leakage

Limitations and Bias

Domain specificity: Trained primarily on religious/biblical Aramaic texts
Vocabulary coverage: Limited to the vocabulary present in the training corpus
Length constraints: Maximum input/output length of 512 tokens
Style consistency: May not handle modern Aramaic dialects or contemporary usage
Performance: Character accuracy of ~64% indicates room for improvement

Environmental Impact

Hardware used: NVIDIA GPU (GTX 3060 12GB)
Training time: ~6 hours
Carbon emissions: Estimated low (single GPU, moderate training time)
Energy efficiency: FP16 mixed precision used to reduce memory usage

Citation

If you use this model in your research, please cite:

bibtex

@misc{aramaic-diacritization-2024,
  title={Aramaic Diacritization Model},
  author={John Locke Jr.},
  year={2025},
  howpublished={Hugging Face Model Hub},
  url={https://huggingface.co/johnlockejrr/aramaic-diacritization-model}
}

License

[MIT]

Acknowledgments

Base model: Helsinki-NLP/opus-mt-afa-afa
Training framework: Hugging Face Transformers
Dataset: Custom Aramaic diacritization corpus

Model Files

model.safetensors - Model weights (234MB)
config.json - Model configuration
tokenizer_config.json - Tokenizer configuration
source.spm / target.spm - SentencePiece models
vocab.json - Vocabulary file
generation_config.json - Generation parameters

Training Scripts

The model was trained using custom scripts:

train_arc2arc_improved_deep.py - Main training script
run_arc2arc_improved_deep.sh - Training execution script
run_resume_arc2arc_deep.sh - Resume training script

Contact

For questions, issues, or contributions, please open an issue on the model repository.