SlangGPT User Feedback Dataset

📖 Dataset Description

This dataset contains human feedback on translations produced by SlangGPT, a model designed to convert Egyptian Arabic dialect into Modern Standard Arabic (MSA). Users evaluate the quality and correctness of generated translations.

🎯 Purpose

This dataset was created to:

• Evaluate SlangGPT's real-world translation quality.
• Build labeled data for training a translation correctness detector.
• Support reinforcement learning from human feedback (RLHF).
• Improve future Egyptian Arabic NLP systems.

🧾 Data Collection Process

Users were shown:

1. An Egyptian Arabic sentence.
2. The generated MSA translation from SlangGPT.

They were then asked:

• Is this translation correct? → Correct / Incorrect
• Rate the quality (required) → 0–5
• Provide a better translation (required if incorrect or rating ≤ 2)

All responses were timestamped and stored.

---

📁 Dataset Fields

---

📊 Example Entry

{
  "egyptian_arabic": "انا زعلان منك جدا",
  "generated_msa": "أنا آسف لك جدا",
  "user_label": "incorrect",
  "user_rating": 0,
  "corrected_msa": "انا حزين منك جدا",
  "timestamp": "2026-05-24T13:24:12.159885Z"
}

---

📈 Dataset Statistics

---

🚀 Usage

Load the Dataset

from datasets import load_dataset

dataset = load_dataset("AdhamAshraf/slanggpt-feedback-dataset")
df = dataset["train"].to_pandas()

Filter by Label

correct_df   = df[df["user_label"] == "correct"]
incorrect_df = df[df["user_label"] == "incorrect"]

Get Correction Pairs for Fine-tuning

# All rows where a human provided a better translation
corrections = df[df["corrected_msa"] != ""]

# Group multiple corrections per sentence
mapping = (
    corrections
    .groupby("egyptian_arabic")["corrected_msa"]
    .apply(list)
    .reset_index()
)

Filter by Quality

high_quality = df[df["user_rating"] >= 4]
low_quality  = df[df["user_rating"] <= 2]

---

🤖 Suggested Use Cases

• Machine Translation Evaluation
• RLHF Reward Modeling
• Translation Quality Classification
• Egyptian Arabic NLP Research
• Human Preference Modeling

---

⚖️ License

This dataset is released under the MIT License. It may be used freely for academic and commercial purposes with attribution.

---

🙏 Acknowledgements

Dataset curated by:

• Adham Ashraf
• Abdelrahman Ahmed
• Ahmed Fekry

---

📚 Related Resources

• GitHub: https://github.com/adhamashraf7788/SlangGPT
• Main Dataset: https://huggingface.co/datasets/AdhamAshraf/egyptian-2-arabic
• Space: https://huggingface.co/spaces/AdhamAshraf/SlangGPT

---

📝 Citation

@dataset{slanggpt_feedback_2026,
  title={SlangGPT User Feedback Dataset},
  author={Abdelrahman Ahmed and Adham Ashraf and Ahmed Fekry},
  year={2026},
  publisher={Hugging Face}
}