ID_REG_Embed_2510 / README.md
Azzindani's picture
Upload README.md with huggingface_hub
a2b18b2 verified
metadata
license: cc-by-4.0
task_categories:
  - text-generation
  - question-answering
  - table-to-text
language:
  - id
tags:
  - legal
  - indonesia
  - knowledge-graph
  - rag
  - production-ready
  - ontology
  - regulation

ID_REG_Embed_2510

This dataset is a production-grade collection of Indonesian regulations, meticulously extracted and processed to support advanced Retrieval-Augmented Generation (RAG) and legal AI workflows. It leverages a hybrid search approach by providing both high-dimensional semantic embeddings and dense keyword vectors.

Dataset Description

The dataset contains approximately 50,000 processed regulation entries. The data was extracted from original PDF sources into structured Markdown format using LLM-based extraction to ensure high fidelity for tables, hierarchies, and legal formatting.

Key Features

  • Real Regulation Data: Sourced from authentic Indonesian regulatory documents.
  • Semantic Embeddings: Includes 1024-dimension embeddings optimized for deep semantic retrieval.
  • Keyword Vectors: Includes 20,000-dimension TF-IDF vectors to support precise keyword matching.
  • Hybrid Search Ready: Designed specifically for systems combining semantic search and traditional BM25/TF-IDF keyword search for maximum accuracy.
  • Structured Format: Data is cleaned and provided in a format suitable for immediate ingestion into vector databases or knowledge graphs.

Development & Methodology

This project was executed under a strict "high-efficiency, limited-budget" framework, demonstrating a cost-effective pipeline for professional legal data engineering:

  • Extraction: Utilized a cloud-to-cloud pipeline with remote access to convert PDF regulations into Markdown. The LLM extraction process was optimized to fit a ~50 USD budget for the initial 50K documents.
  • Vector Generation: Embeddings and TF-IDF matrices were generated by leveraging a combination of free-tier hybrid instances and paid cloud GPU/CPU compute (approx. 10 USD total cost).
  • Weaknesses: Due to budget constraints, the current version is limited to the first 50,000 regulations. Future iterations may expand this scope as resources allow.

Technical Specifications

Feature Specification
Language Indonesian (ID)
Embedding Dimensions 1024
TF-IDF Dimensions 20,000
Format Parquet / JSONL (Structured)
Total Rows ~50,000

Use Cases

  1. Legal RAG Systems: Building chatbots or research tools that require grounding in Indonesian law.
  2. Ontology Building: Extracting entities and relationships for legal knowledge graphs.
  3. Semantic Search: Improving document retrieval beyond simple keyword matching.
  4. Legal Analysis: Text mining and trend analysis within the Indonesian regulatory landscape.

License

This dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). You are free to share and adapt the material for any purpose, even commercially, as long as appropriate credit is given.