metadata

library_name: transformers
license: cc-by-nc-4.0
base_model: atlasia/Terjman-Large-v1.2
metrics:
  - bleu
  - chrf
  - ter
model-index:
  - name: Terjman-Large-v2.0
    results: []
datasets:
  - BounharAbdelaziz/Terjman-v2-English-Darija-Dataset-350K
language:
  - ary
  - en
pipeline_tag: translation

🇲🇦 Terjman-Large-v2.0 (240M) 🚀

Terjman-Large-v2.0 is an improved version of atlasia/Terjman-Large-v1.2, built on the powerful Transformer architecture and fine-tuned for high-quality, accurate translations.

This version is based on atlasia/Terjman-Large-v1.2 and has been trained on a larger and more refined dataset, leading to improved translation performance. The model achieves results on par with gpt-4o-2024-08-06 on TerjamaBench, an evaluation benchmark for English-Moroccan darija translation models, that challenges the models more on the cultural aspect.

🚀 Features

✅ Fine-tuned for English->Moroccan darija translation.
✅ State-of-the-art performance among open-source models.
✅ Compatible with 🤗 Transformers and easily deployable on various hardware setups.

🔥 Performance Comparison

The following table compares Terjman-Large-v2.0 against proprietary and open-source models using BLEU, chrF, and TER scores. Higher BLEU/chrF and lower TER indicate better translation quality.

Model	Size	BLEU↑	chrF↑	TER↓
Proprietary Models
gemini-exp-1206	*	30.69	54.16	67.62
claude-3-5-sonnet-20241022	*	30.51	51.80	67.42
gpt-4o-2024-08-06	*	28.30	50.13	71.77
Open-Source Models
Terjman-Ultra-v2.0	1.3B	25.00	44.70	77.20
Terjman-Supreme-v2.0	3.3B	23.43	44.57	78.17
Terjman-Large-v2.0 (This model)	240M	22.67	42.57	83.00
Terjman-Nano-v2.0	77M	18.84	38.41	94.73
atlasia/Terjman-Large-v1.2.2	240M	16.33	37.10	89.13
MBZUAI-Paris/Atlas-Chat-9B	9B	14.80	35.26	93.95
facebook/nllb-200-3.3B	3.3B	14.76	34.17	94.33
atlasia/Terjman-Nano	77M	09.98	26.55	106.49

🔬 Model Details

Base Model: atlasia/Terjman-Large-v1.2
Architecture: Transformer-based sequence-to-sequence model
Training Data: High-quality parallel corpora with high quality translations
Training Precision: FP16 for efficient inference

🚀 How to Use

You can use the model with the Hugging Face Transformers library:

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_name = "BounharAbdelaziz/Terjman-Large-v2.0"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

def translate(text):
    inputs = tokenizer(text, return_tensors="pt")
    output = model.generate(**inputs)
    return tokenizer.decode(output[0], skip_special_tokens=True)

# Example translation
text = "Hello there! Today the weather is so nice in Geneva, couldn't ask for more to enjoy the holidays :)"
translation = translate(text)
print("Translation:", translation)
# prints: صباح الخير! اليوم الطقس زوين بزاف فجنيف، ما قدرتش نطلب المزيد باش نستمتعو بالعطل:)

🖥️ Deployment

Run in a Hugging Face Space

Try the model interactively in the Terjman-Large Space 🤗

Use with Text Generation Inference (TGI)

For fast inference, use Hugging Face TGI:

pip install text-generation
text-generation-launcher --model-id BounharAbdelaziz/Terjman-Large-v2.0

Run Locally with Transformers & PyTorch

pip install transformers torch
python -c "from transformers import pipeline; print(pipeline('translation', model='BounharAbdelaziz/Terjman-Large-v2.0')('Hello there!'))"

Deploy on an API Server

Use FastAPI to serve translations as an API:

from fastapi import FastAPI
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

app = FastAPI()
model_name = "BounharAbdelaziz/Terjman-Large-v2.0"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

@app.get("/translate/")
def translate(text: str):
    inputs = tokenizer(text, return_tensors="pt")
    output = model.generate(**inputs)
    return {"translation": tokenizer.decode(output[0], skip_special_tokens=True)}

🛠️ Training Details Hyperparameters**

The model was fine-tuned using the following training settings:

Learning Rate: 0.001
Training Batch Size: 16
Evaluation Batch Size: 16
Seed: 42
Gradient Accumulation Steps: 8
Total Effective Batch Size: 128
Optimizer: AdamW (Torch) with betas=(0.9,0.999), epsilon=1e-08
Learning Rate Scheduler: Linear
Warmup Ratio: 0.1
Epochs: 2
Precision: Mixed FP16 for efficient training

Framework versions

Transformers 4.47.1
Pytorch 2.5.1+cu124
Datasets 3.1.0
Tokenizers 0.21.0

📜 License

This model is released under the CC BY-NC (Creative Commons Attribution-NonCommercial) license, meaning it can be used for research and personal projects but not for commercial purposes. For commercial use, please get in touch :)

@misc{terjman-v2,
  title = {Terjman-v2: High-Quality English-Moroccan Darija Translation Model},
  author={Abdelaziz Bounhar},
  year={2025},
  howpublished = {\url{https://huggingface.co/BounharAbdelaziz/Terjman-Large-v2.0}},
  license = {CC BY-NC}
}