epdk_corpus / README.md
ogulcanakca's picture
Update README.md
f305eef verified
metadata
pretty_name: EPDK Corpus (v2)
tags:
  - turkish
  - energy-market
  - legislation
  - regulations
  - epdk
  - ocr
  - qwen-vl
  - text-extraction
  - domain-specific
  - language-modeling
  - legal
dataset_info:
  features:
    - name: id
      dtype: string
    - name: text
      dtype: string
    - name: kind
      dtype: string
    - name: method
      dtype: string
  splits:
    - name: train
      num_examples: 3577
  download_size: 72000000
  dataset_size: 145000000
configs:
  - config_name: v2
    data_files:
      - split: train
        path: epdk_corpus_dataset.jsonl
language:
  - tr
license: apache-2.0
task_categories:
  - text-generation
size_categories:
  - 1K<n<10K
version: 2.0.0

Dataset Card for EPDK Electricity Market Corpus (v2)

Dataset Summary

This corpus contains high-quality Turkish text extracted from documents published by the Turkish Energy Market Regulatory Authority (EPDK – Enerji Piyasası Düzenleme Kurumu).
The source materials include regulations, board decisions, and official communiqués related to the electricity market legislation.

Version 2.0.0 introduces a complete re-extraction of the text using the Qwen-2.5-VL vision-language model, which achieves dramatically better OCR accuracy than Tesseract—particularly on scanned and complex multi-column PDFs.
The result is a significantly cleaner and more semantically faithful representation of the original legal texts.


Version History

Version Description Date Notes
v1.0.0 Initial release using Tesseract OCR (≈ 3.4 K docs) 2025-04 Baseline version. Stored under /v1.0.0/ branch.
v2.0.0 Re-extracted with Qwen-2.5-VL OCR (≈ 3.6 K docs), normalized fields (id, text, kind, method) 2025-10 Recommended for downstream use.

To load a specific version:

from datasets import load_dataset
ds = load_dataset("ogulcanakca/epdk_corpus", revision="v2.0.0")

Intended Uses

Designed primarily for language-model domain adaptation and continued pre-training on Turkish regulatory language. Potential downstream tasks include:

  • Legal Question Answering over EPDK documents
  • Summarization of laws and regulations
  • Document classification or clustering (e.g., tebliğ / karar / yönetmelik)
  • Information retrieval and RAG training for energy-sector LLMs

Languages

All text is in Turkish (tr).


Data Fields

Field Type Description
id string Unique SHA-1-based identifier for the document
text string Cleaned text extracted from the source file
kind string File type (docx, pdf, xls, etc.)
method string Extraction method (python-docx, pandas, qwen-vl-ocr, etc.)

Example Instance

{
  "id": "b3f7a6d9c49e",
  "text": "Enerji Piyasası Düzenleme Kurumundan:\nKURUL KARARI\nKarar No: 9284 Karar Tarihi: 02/04/2020\n...",
  "kind": "docx",
  "method": "python-docx"
}

Data Structure & Access

  • Format: JSONL (one document per line)
  • Total examples: ≈ 3 577
  • Dataset path: v2/final_corpus_dataset.jsonl

@dataset{ogulcanakca_epdk_corpus_v2,
  author    = {Oğulcan Akca},
  title     = {EPDK Corpus (v2)},
  year      = {2025},
  url       = {https://huggingface.co/datasets/ogulcanakca/epdk_corpus}
}

akca_ogulcan@hotmail.com