README.md

metadata

dataset_info:
  - config_name: corpus
    features:
      - name: corpus-id
        dtype: string
      - name: image
        dtype: image
      - name: category
        dtype: string
    splits:
      - name: test
        num_bytes: 221322978.597
        num_examples: 1101
    download_size: 215590614
    dataset_size: 221322978.597
  - config_name: qrels
    features:
      - name: query-id
        dtype: int64
      - name: corpus-id
        dtype: string
      - name: score
        dtype: int64
    splits:
      - name: test
        num_bytes: 67500
        num_examples: 1500
    download_size: 22280
    dataset_size: 67500
  - config_name: queries
    features:
      - name: query-id
        dtype: int64
      - name: query
        dtype: string
    splits:
      - name: test
        num_bytes: 131061
        num_examples: 1500
    download_size: 78796
    dataset_size: 131061
configs:
  - config_name: corpus
    data_files:
      - split: test
        path: corpus/test-*
  - config_name: qrels
    data_files:
      - split: test
        path: qrels/test-*
  - config_name: queries
    data_files:
      - split: test
        path: queries/test-*

KoViDoRe - VQA v1.0

This dataset is part of the KoViDoRe Benchmark for Korean visual document retrieval evaluation. Specifically, it is focused on Visual Question Answering (VQA) tasks for Korean document images.

Dataset Summary

• Domain: Korean structured document images (e.g., 환경기상, 공공행정 등 다양한 카테고리)
• Task: Visual document retrieval (text question → relevant document image)
• Format: BEIR-compatible (corpus, queries, qrels)

Statistics:

• Number of Documents (Pages): 1,100+
• Number of Queries: 1,500
• Number of Relevance Judgments (qrels): 1,500
• All text queries are in Korean.

Dataset Structure

• corpus
  • corpus-id: Unique page ID
  • image: Document page image (PIL Image)
  • category: Document category label (e.g., 환경기상, 공공행정 등)
• queries
  • query-id: Unique query ID
  • query: Korean text question
• qrels
  • query-id
  • corpus-id
  • score: (1 = relevant)

Usage

This dataset is intended for benchmarking the performance of visual retrieval systems on on real-world Korean document images.

Note:
Due to licensing restrictions, original images are not redistributed with this dataset.
Please download the required image data directly from AI Hub (시각화 자료 질의응답 데이터) and place them as instructed in the KoVidore-benchmark GitHub repository.

Acknowledgments

• Based on data provided by AI Hub (시각화 자료 질의응답 데이터)

---