Hedi-Bk/Synthetic_Tunisian_Legal_Dataset

llm-distillation , knowledge-distillation , synthetic-data , legal-ai ,tunisian-law , qlora fine-tuning , rag arabic-nlp , french-nlp # **Legal Dataset Generation with a Teacher LLM** This project is dedicated to creating a specialized legal dataset for Tunisian law. The generation process is assisted by a "Teacher LLM" to produce complex and relevant question-answer examples. The primary tool used for this generation is **Google AI Studio** with the **Gemini 3 Pro** model. ## **Repository Structure** This repository contains the following elements: - **/Batch_Outputs/**: This folder contains all generated data batches in JSON format. Each file represents a batch of question-answer pairs based on a specific legal text. - **Output_Sample.json**: This file serves as a template and example for the expected JSON structure for each data batch. ## **Data Structure** Each JSON file in `Batch_Outputs` follows a precise structure to ensure dataset consistency. Here are the main fields: - `batch_id`: A unique identifier for the batch. - `source`: The name of the source legal text (e.g., "Tunisian Customs Code"). - `articles_covered`: A list of article numbers used to generate the examples in this batch. - `examples`: A list of objects, where each object is a question-answer example. - `example_id`: A unique identifier for the example. - `language`: The language of the example (`fr`, `ar`, `derja`). - `category`: The category of legal reasoning (e.g., `direct`, `multi_article`, `principle_exception`). - `messages`: A structured conversation between a `system`, a `user`, and an `assistant`. - `coverage_summary`: A summary of the article coverage from the legal text within the batch. ## **Contents of Data** - Multi-hop reasoning QA pairs - Context provided with each question (to simulate chunks coming from RAG to LLMs + Query) - Samples in Arabic (MSA and Derja) and French ## **📈 Next Steps** - Scale to 1000+ samples - Human validation with legal expert ## **Keywords** - Legal Tech - Legal Dataset - Data Generation - Large Language Model (LLM) - Gemini - Google AI Studio - Tunisian Law - Question …