samfatnassi/Tunisian-Railway-Dialogues

🇹🇳 Tunisian Railway Dialogues (Cleaned & Verified by Kilma.ai)

🚀 Dataset Summary

This dataset contains 1,720 high-quality spoken dialogues in Tunisian Dialect (Darija), focused on railway services and customer support scenarios.

It is a processed and verified version of the original TuDiCOI corpus, optimized by Kilma.ai for training Large Language Models (LLMs) and Fine-tuning Chatbots.

✅ Kilma.ai Verification Process

At Kilma.ai, we focus on production-ready AI data. We have processed the raw academic XML files through a rigorous cleaning pipeline:

• Filtering: Removed incomplete, empty, or low-quality turns, reducing the original count from 1825 to a curated 1,720 high-quality dialogues.
• Formatting: Converted from complex XML to standard JSONL (Chat Format) compatible with OpenAI, Llama 3, and Mistral fine-tuning.
• Role Mapping: Standardized speaker roles to user (Client) and assistant (Staff/Agent).
• Normalization: Cleaned text from excessive whitespace and non-textual artifacts.

📊 Dataset Statistics

• Total Dialogues: ~1,720
• Language: Tunisian Dialect (Derja) / Arabic Script.
• Domain: Customer Service, Transportation, Railway (SNCFT).
• Tasks: Intent Detection, Slot Filling (Time, Price, Destination), Dialogue Systems.

💡 Use Cases

This dataset is ideal for:

1. Fine-tuning LLMs to understand and speak Tunisian Dialect.
2. Building Customer Support Agents for the North African market.
3. Research in Spoken Language Understanding (SLU) for under-resourced dialects.

📂 Data Structure (JSONL)

Each line in the dataset is a JSON object representing a full conversation:

{
  "messages": [
    {
      "role": "system", 
      "content": "You are a helpful assistant for the Tunisian National Railway Company (SNCFT)."
    },
    {
      "role": "user", 
      "content": "وقتاش يخرج التران لصفاقس؟"
    },
    {
      "role": "assistant", 
      "content": "فما واحد يخرج الماضي ساعة وربع."
    }
  ]
}

📜 Original Source & Citation
This dataset is derived from the TuDiCOI (Tunisian Dialect Corpus Interlocutor) project, originally constructed by the ANLP-RG (Arabic Natural Language Processing Research Group) in cooperation with SNCFT.
If you use this dataset in academic work, please cite the original authors:
    ⁠M. Graja, M. Jaoua, L. Hadrich Belguith, "Statistical Framework with Knowledge Base Integration for Robust Speech Understanding of the Tunisian Dialect", IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 23, NO. 12, DECEMBER 2015.

Maintained by: Kilma.ai - Building the future of Voice AI