Omartificial-Intelligence-Space/Shami-MT

SHAMI-MT : A Machine Translation Model From MSA to Syrian Dialect

This model is based on the paper SHAMI-MT: A Syrian Arabic Dialect to Modern Standard Arabic Bidirectional Machine Translation System.

Model Description

SHAMI-MT is a specialized machine translation model designed to translate from Modern Standard Arabic (MSA) to Syrian dialect. Built on the robust AraT5v2-base-1024 architecture, this model bridges the gap between formal Arabic and the rich dialectal variations of Syrian Arabic.

Model Details

• —Model Type: Sequence-to-Sequence Translation
• —Base Model: UBC-NLP/AraT5v2-base-1024
• —Language: Arabic (MSA → Syrian Dialect)
• —License: Apache 2.0
• —Library: Transformers

Dataset

The model was trained on the Nâbra dataset, a comprehensive corpus of Syrian Arabic dialects with morphological annotations.

Nâbra Dataset Details

Citation:

Nayouf, A., Hammouda, T., Jarrar, M., Zaraket, F., & Kurdy, M. B. (2023). 
Nâbra: Syrian Arabic dialects with morphological annotations. 
arXiv preprint arXiv:2310.17315.

Key Statistics:

• —Tokens: ~60,000 words
• —Dialects Covered: Multiple Syrian regional dialects including:
• —Aleppo
• —Damascus
• —Deir-ezzur
• —Hama
• —Homs
• —Huran
• —Latakia
• —Mardin
• —Raqqah
• —Suwayda

Data Sources:

• —Social media posts
• —Movie and TV series scripts
• —Song lyrics
• —Local proverbs

Training Details

The model was fine-tuned on the AraT5v2-base-1024 architecture with the following training metrics:

• —Total Training Steps: 10,384
• —Epochs: 22
• —Final Training Loss: 1.396
• —Final Evaluation Loss: 0.771
• —Learning Rate: Cosine schedule starting at 5e-5
• —Batch Size: 256
• —Total FLOPs: 1.58e+17

Training Progress

The model showed consistent improvement throughout training:

• —Initial loss: 12.93 → Final loss: 1.40
• —Evaluation loss steadily decreased from 1.44 to 0.77
• —Gradient norms remained stable throughout training

Usage

Installation

pip install transformers torch

Inference Code

from transformers import T5Tokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
tokenizer = T5Tokenizer.from_pretrained("Omartificial-Intelligence-Space/Shami-MT")
model = AutoModelForSeq2SeqLM.from_pretrained("Omartificial-Intelligence-Space/Shami-MT")

# Example usage
ar_prompt = "مرحبا بك هنا"  # MSA input
input_ids = tokenizer(ar_prompt, return_tensors="pt").input_ids
outputs = model.generate(input_ids)

print("Input (MSA):", ar_prompt)
print("Tokenized input:", tokenizer.tokenize(ar_prompt))
print("Output (Syrian Dialect):", tokenizer.decode(outputs[0], skip_special_tokens=True))

Generation Parameters

For optimal results, you can adjust generation parameters:

outputs = model.generate(
    input_ids,
    max_length=128,
    num_beams=4,
    temperature=0.7,
    do_sample=True,
    pad_token_id=tokenizer.pad_token_id,
    eos_token_id=tokenizer.eos_token_id
)

Evaluation Results

• —Test Set: 1,500 unseen sentences
• —Evaluation Method: GPT-4.1 as automated judge
• —Average Score: 4.01/5.0 ⭐
• —Evaluation Criteria: Translation quality, dialectal accuracy, and semantic preservation

The model was evaluated using GPT-4.1 as an automated judge with the following structured prompt:

"You are a language evaluation assistant. Compare the predicted Shami sentence to the reference.
Please return a rating from 0 to 5 and a short comment.

MSA Input: [input sentence]
Model Prediction (Shami dialect): [model output]
Ground Truth (Shami dialect): [reference translation]

Respond in this format:
Score: <number from 0 to 5>
Comment: <brief explanation of the score>"

Score Distribution Analysis:

• —Excellent (5.0): High-quality translations with perfect dialectal conversion
• —Good (4.0-4.9): Minor dialectal variations or stylistic differences
• —Average (3.0-3.9): Acceptable translations with some dialectal inconsistencies
• —Below Average (2.0-2.9): Noticeable errors in dialect or meaning
• —Poor (0-1.9): Significant translation errors or loss of meaning

Performance Highlights

• —Strong Dialectal Conversion: Successfully transforms MSA into authentic Syrian dialect
• —Semantic Preservation: Maintains original meaning while adapting linguistic style
• —Regional Adaptability: Handles various Syrian sub-dialects effectively
• —Consistent Quality: Stable performance across different text types and domains

Applications

This model is particularly useful for:

• —Content Localization: Adapting MSA content for Syrian audiences
• —Cultural Preservation: Maintaining and promoting Syrian dialectal variations
• —Educational Tools: Teaching differences between MSA and Syrian dialect
• —Research: Syrian Arabic NLP and dialectology studies

Regional Coverage

The model handles multiple Syrian sub-dialects, making it versatile for different regions within Syria:

🏛️ Urban Centers: Damascus, Aleppo 🏔️ Northern Regions: Latakia, Mardin 🏜️ Eastern Areas: Deir-ezzur, Raqqah 🌄 Central/Southern: Hama, Homs, Huran, Suwayda

Limitations

• —Trained specifically on Syrian dialect variations
• —Performance may vary for other Arabic dialects
• —Limited to text-based translation (no speech support)
• —Dataset size constraints may affect handling of very rare dialectal expressions

Citation

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

@misc{shami-mt-2024,
  title={SHAMI-MT: A Machine Translation Model From MSA to Syrian Dialect},
  author={Omartificial Intelligence Space},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/Omartificial-Intelligence-Space/Shami-MT}
}

@article{nayouf2023nabra,
  title={Nâbra: Syrian Arabic dialects with morphological annotations},
  author={Nayouf, Amal and Hammouda, Tymaa Hasanain and Jarrar, Mustafa and Zaraket, Fadi A and Kurdy, Mohamad-Bassam},
  journal={arXiv preprint arXiv:2310.17315},
  year={2023}
}

@misc{onajar2025shamiMT,
  title={Shami-MT-2MSA : A Machine Translation from Syrian Dialect to MSA},
  author={Sibaee, Serry and Nacar, Omer},
  year={2025}
}

Contact & Support

For questions, issues, or contributions, please visit the model repository or contact the development team.