KoSTA-dataset / README.md
anonymous128463's picture
Upload folder using huggingface_hub
821c929 verified
|
Raw
History Blame Contribute Delete
1.77 kB
metadata
license: cc-by-nc-4.0
language:
  - ko
task_categories:
  - image-to-image
tags:
  - scene-text-editing
  - korean
  - pii
  - ocr
  - benchmark
size_categories:
  - n<1K

KoSTA-bench

한국어 PII(개인정보) 장면 텍스트 편집(Scene Text Editing) 벤치마크. 원본·타깃 텍스트를 모두 가상 PII 키워드로 구성한 611개 (원본, 타깃) 이미지 쌍으로 이루어진다.

구성

  • source/{name} — 원본 텍스트(text_s)가 담긴 원본 이미지 crop
  • target/{name} — 타깃 텍스트(text_t)로 편집된 타깃 이미지 crop
  • manifest.jsonl — 샘플별 메타데이터 (한 줄당 한 샘플)

manifest 필드

필드 설명
name 샘플 파일명 (source/, target/ 공통)
text_s 원본 텍스트
text_t 타깃 텍스트
label PII 유형 라벨

예시:

{"name": "000000.png", "text_s": "원수", "text_t": "과장", "label": "cv_position"}

통계

10개 카테고리 21종 유형, 611쌍(원본·타깃 합 1,222장). 모든 샘플은 PaddleOCR PP-OCRv5 완전일치(strict-pass) 검증과 전문가 3인 중 2인 이상의 인간 검수를 통과했다.

로딩 예시

import json
from huggingface_hub import snapshot_download

root = snapshot_download("<your-username>/KoSTA-bench", repo_type="dataset")
rows = [json.loads(l) for l in open(f"{root}/manifest.jsonl")]

r = rows[0]
src_path = f"{root}/source/{r['name']}"   # 원본 이미지
tgt_path = f"{root}/target/{r['name']}"   # 타깃 이미지

라이선스 및 개인정보

원본·타깃 텍스트는 모두 가상 PII 키워드이며 실제 개인정보를 포함하지 않는다. license 필드는 실제 배포 근거에 맞게 수정하여 사용한다.