| language: | |
| - uz | |
| - ru | |
| task_categories: | |
| - text-generation | |
| - question-answering | |
| pretty_name: Uzbek Legal Dataset (LexUZ) | |
| size_categories: | |
| - 1K<n<10K | |
| license: mit | |
| # Uzbek Legal Dataset (LexUZ) | |
| ## Description | |
| This dataset contains cleaned legal documents from LexUZ, prepared for instruction fine-tuning of large language models (LLMs). | |
| It includes structured legal metadata and ChatML-formatted instruction-response pairs. | |
| --- | |
| ## Dataset Structure | |
| ### 1. Instruction Dataset | |
| - `train.jsonl` | |
| - Format: ChatML | |
| - Used for fine-tuning | |
| ### 2. Structured Dataset | |
| - `documents.csv` | |
| - Tabular format for analysis / preprocessing | |
| --- | |
| ## Features | |
| Each document includes: | |
| - Document title | |
| - Document type (law, decree, resolution) | |
| - Issuer | |
| - Document number | |
| - Date | |
| - Cleaned legal text | |
| --- | |
| ## Format | |
| Each training sample is structured as: | |
| ```json | |
| { | |
| "messages": [ | |
| {"role": "system", "content": "..."}, | |
| {"role": "user", "content": "..."}, | |
| {"role": "assistant", "content": "..."} | |
| ] | |
| } |