SlangGPT

Egyptian Arabic → Modern Standard Arabic

Fine-tuning AraGPT-2 for dialect-to-MSA generation and translation detection

Resource	Link
🤗 Live Demo	SlangGPT Space
📄 Paper	SlangGPT Report
📦 Training Dataset	egyptian-2-arabic
💬 Feedback Dataset	slanggpt-feedback-dataset
📦 Kaggle	egyptian-2-arabic

---

Motivation

Over 100 million Egyptians speak Egyptian Arabic daily — yet most Arabic NLP systems are trained almost entirely on Modern Standard Arabic (MSA). Egyptian Arabic is not simply a simplified form of MSA; it carries distinct vocabulary, grammar, and culturally loaded expressions that MSA-trained models consistently fail to handle. Words like جدع carry meanings of honor, loyalty, and social standing that have no direct MSA equivalent, yet they appear constantly in everyday Egyptian speech.

When we set out to build SlangGPT, we discovered there was no large, clean, publicly available Egyptian Arabic ↔ MSA parallel dataset suitable for this task. So we built one ourselves.

---

Overview

SlangGPT fine-tunes AraGPT-2 on an 18,250-sentence parallel Egyptian Arabic / MSA corpus to solve two tasks:

• Generation — Given an Egyptian Arabic sentence, generate the equivalent MSA translation
• Detection — Given an (Egyptian, MSA) pair, classify whether the translation is correct

This project extends the methodology of Hernandez & Naik (Stanford CS224N, 2025) — who fine-tuned GPT-2 for Gen-Z English slang understanding — to the Arabic dialect setting, replacing the English backbone with AraGPT-2 and the slang dataset with a parallel Egyptian–MSA corpus.

---

Results

Task	Model	chrF	BLEU	Accuracy
Generation	Zero-shot AraGPT-2	10.62	0.02	—
Generation	Fine-tuned AraGPT-2	29.08	6.63	—
Detection	Zero-shot AraGPT-2	—	—	0.500
Detection	Fine-tuned AraGPT-2	—	—	0.956

Fine-tuning improves detection accuracy by +45.6 points and generation chrF by +18.5 points over zero-shot baselines.

Generation Examples

Input (Egyptian)	Zero-shot Output	Fine-tuned Output
يلا فين؟	مالذي جاء به من خير... (forum drift)	هيا، أين أنت؟
أنا محتاج أتكلم معاكي	ياام.. يآآإك ياروحيتي... (social media drift)	أحتاج أن أتحدث معك
كنت فاكرك مش جاية	يااللي ما انخطبتك... (forum content)	كنت أذكرك، لستِ قادمة

---

Known Limitations

• Generation quality is still limited. 18K sentence pairs is relatively small for a full translation task. BLEU of 6.63, while a significant improvement over zero-shot, indicates the model is not yet production-ready for translation.
• AraGPT-2 was not pretrained on dialectal Arabic. The backbone was trained on MSA and online Arabic text, which creates an inherent ceiling on how well it can generate fluent dialectal-to-MSA translations without a much larger fine-tuning corpus.
• Culturally loaded words are hard to translate. Terms like جدع, أوي, or خلاص carry meanings that are difficult to capture in MSA even for humans. The model often produces technically correct but culturally flat translations.
• Dataset coverage is uneven. The 18K corpus skews toward conversational sentences and may not generalize well to domain-specific Egyptian Arabic (e.g., medical, legal, or technical speech).

We deployed the demo publicly with a community feedback loop specifically to address these limitations over time.

---

Community Feedback & Retraining

The live demo includes a feedback system where users can rate translations and submit corrections. Every submission is automatically saved to the open feedback dataset and will be used for periodic retraining to improve generation quality continuously.

from datasets import load_dataset
df = load_dataset("AdhamAshraf/slanggpt-feedback-dataset", split="train").to_pandas()

# High quality confirmed translations
high_quality = df[df["user_rating"] >= 4]

# Human corrections for fine-tuning
corrections = df[df["corrected_msa"] != ""]

---

Datasets

Training Dataset — egyptian-2-arabic

18,250 parallel Egyptian Arabic / MSA sentence pairs used to train both the generation and detection models.

Split	Generation Pairs	Detection Examples
Train (80%)	14,600	29,200
Dev (10%)	1,825	3,650
Test (10%)	1,825	3,650

🔗 huggingface.co/datasets/AdhamAshraf/egyptian-2-arabic

from datasets import load_dataset
dataset = load_dataset("AdhamAshraf/egyptian-2-arabic", split="train")
df = dataset.to_pandas()

Source & Derivation

Derived from Abdalrahmankamel/Egyptian_2_English. Modifications:

• Removed English translation column
• Added Modern Standard Arabic translations
• Applied Arabic normalization and diacritic removal
• Reformatted for dialect-to-MSA tasks

Feedback Dataset — slanggpt-feedback-dataset

Human feedback collected from the live Space. Users rate model translations and provide corrections, forming a growing dataset for future fine-tuning.

Field	Type	Description
egyptian_arabic	string	Original Egyptian Arabic input
generated_msa	string	SlangGPT's generated translation
user_label	string	correct or incorrect
user_rating	int64	Quality score 0–5
corrected_msa	string	Human correction (required if incorrect or rating ≤ 2)
timestamp	string	ISO 8601 UTC timestamp

---

Quickstart

Note: GPU required. CPU inference is too slow for practical use. Training was done on a T4 GPU via Google Colab.

1. Clone and install

git clone https://github.com/adhamashraf7788/SlangGPT.git
cd SlangGPT
pip install -r requirements.txt

2. Download model weights

python scripts/download_weights.py

Downloads weights to:

model/weights/detection/best_model.pt            (~527 MB)
model/weights/generation/best/model.safetensors  (~1.37 GB)

3. Download and preprocess the dataset

from datasets import load_dataset
dataset = load_dataset("AdhamAshraf/egyptian-2-arabic", split="train")
df = dataset.to_pandas()
df.to_csv("data/raw/NLP.csv", index=False, encoding="utf-8-sig")

python data/prepare_data.py --raw_csv data/raw/NLP.csv

4. Run the web app

python app/app.py

Open http://localhost:5000

---

Training

Training was done on Google Colab (T4 GPU). Open the notebooks in order:

Notebook	Description
01_preprocessing.ipynb	Download dataset, clean, split, build detection pairs
02_train_generation.ipynb	Fine-tune AraGPT-2 medium for generation
03_train_detection.ipynb	Fine-tune AraGPT-2 base for detection

Hyperparameters

	Generation	Detection
Base model	aragpt2-medium	aragpt2-base
Parameters	~355M	~135M
Learning rate	5e-5	2e-5
Batch size	8 (eff. 32)	16
LR schedule	Cosine	Linear
Warmup ratio	10%	10%
Weight decay	0.01	0.01
Epochs	5 (best at ep. 3)	8
Train loss (start → end)	2.50 → 0.76	0.71 → 0.10

---

Evaluation

# Zero-shot baseline
python evaluation/baseline.py

# Fine-tuned model evaluation (chrF, BLEU, PPL, accuracy)
python evaluation/evaluate.py

# Error analysis (false positives / false negatives)
python evaluation/error_analysis.py

Detection Error Breakdown (Test Set)

	Count	Rate
Total test examples	3,650	—
Correct predictions	3,491	95.6%
False Positives	101	2.8%
False Negatives	58	1.6%

---

Models

Task	Base Model	Link
Generation	AraGPT-2 Medium	aubmindlab/aragpt2-medium
Detection	AraGPT-2 Base	aubmindlab/aragpt2-base

Generation uses causal language modeling with prompt masking — only the MSA target tokens contribute to the training loss. Inference uses greedy decoding with repetition penalty.

Detection encodes a cloze-style Arabic prompt through AraGPT-2 and passes the last-token hidden state through a linear classifier head trained with binary cross-entropy.

---

Project Structure

SlangGPT/
├── app/                          # Flask web application
│   ├── app.py
│   ├── model.py
│   ├── templates/index.html
│   └── static/style.css
├── data/
│   ├── prepare_data.py
│   ├── raw/NLP.csv               # [git-ignored]
│   └── processed/                # [git-ignored]
├── model/
│   ├── config.py
│   ├── train_generation.py
│   ├── train_detection.py
│   └── weights/                  # [git-ignored]
├── evaluation/
│   ├── baseline.py
│   ├── evaluate.py
│   ├── error_analysis.py
│   └── plots/
├── notebooks/                    # [git-ignored]
├── scripts/
│   └── download_weights.py
├── report/
│   ├── main.tex
│   └── references.bib
├── requirements.txt
└── .gitignore

---

Related Work

This project extends:

Hernandez & Naik, Extending GPT-2 for Informal and Slang Aware Language Understanding, Stanford CS224N, 2025

Which builds on:

• Antoun et al., AraGPT2, 2021
• Sun et al., Toward Informal Language Processing, 2024
• Radford et al., GPT-2, 2019

---

Citation

@misc{slanggpt2026,
  title={SlangGPT: Fine-tuning AraGPT-2 for Egyptian Arabic to MSA Generation and Detection},
  author={Abdelrahman Ahmed and Adham Ashraf and Ahmed Fekry},
  year={2026},
  url={https://github.com/adhamashraf7788/SlangGPT}
}

---

License

This project is licensed under the MIT License — see the LICENSE file for details.