You can deploy this model and start earning money today!

faresalawneh/
jordanian-to-fusha-model

Name: jordanian-to-fusha-model
Availability: InStock

facebook/nllb-200-distilled-600M

finetune

mit

Views

No views yet

🇯🇴 Jordanian Dialect → Classical Arabic (فصحى)

A fine-tuned sequence-to-sequence model that translates Jordanian Arabic dialect (عامية أردنية) into Classical Arabic / Modern Standard Arabic (فصحى), built on top of Facebook's NLLB-200-distilled-600M.

Model Details

Developed by: Fares Alawneh
Model type: Seq2Seq Translation (fine-tuned NLLB)
Language(s): Arabic — Levantine dialect → Modern Standard Arabic
License: MIT
Finetuned from: facebook/nllb-200-distilled-600M
Dataset: Jordanian Dialect Dataset

Performance

Metric	Score
BLEU Score (Validation)	41.6
Training Samples	~30,000
Training Epochs	2

How to Use

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("faresalawneh/jordanian-to-fusha-model")
tokenizer = AutoTokenizer.from_pretrained("faresalawneh/jordanian-to-fusha-model")

def translate(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=96)
    outputs = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids("arb_Arab"),
        max_length=96,
        num_beams=4
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

print(translate("شو عم تعمل هلق؟"))

Training Details

Parameter	Value
Base Model	`facebook/nllb-200-distilled-600M`
Source Language	`apc_Arab` (Levantine Arabic)
Target Language	`arb_Arab` (Modern Standard Arabic)
Max Input Length	96 tokens
Max Output Length	96 tokens
Batch Size	32
Learning Rate	5e-4
Epochs	2
Hardware	Google Colab T4 GPU

Training Data

The model was trained on a Jordanian dialect dataset containing ~30,000 comment pairs with the following columns:

Text — original Jordanian dialect text
Text Corrected — Classical Arabic translation
Has Error — rows with Has Error == 1 were excluded

Dataset available on Kaggle: Jordanian Dialect Dataset

Limitations

Primarily trained on Jordanian dialect — may not generalize well to other Levantine dialects (Syrian, Lebanese, Palestinian)
Informal slang and very colloquial expressions may not translate accurately
Performance degrades on very long sentences (>96 tokens)

Author

Fares Alawneh AI Student — Jordan University of Science and Technology GitHub · Kaggle