andynoodles/Taiwan-JudicialYuanPublication

司法周刊 Judicial Weekly OCR — block-level (zh-Hant, vertical text)

Block-level OCR pairs synthesised from scanned issues of 司法周刊 (Judicial Weekly), the official weekly newspaper of Taiwan's Judicial Yuan (司法院). Each row is one cropped layout region with its block type and the vision-LLM transcription.

• —104,938 rows from 1,506 scanned pages (one PDF per 版-group, 期1–756)
• —Complete scan era 1981–1995 (民國70–84), ~100 issues per year, every year covered
• —Vertical Traditional Chinese (直書): columns read top-to-bottom, right-to-left — rare, high-value layout for OCR training
• —Source: 司法院全球資訊網 (www.judicial.gov.tw) public archive — JBIG2 bilevel letterpress scans, 100% scanned (no born-digital pages)

How it was made

Scanned page → layout detection (PP-DocLayoutV3) → per-region crops → OCR by a PaddleOCR-VL vision-LLM (the teacher) → revision by a Qwen3.6 vision-LLM (the reviser) → Traditional-Chinese normalization (OpenCC s2t). This is a distillation dataset: there is no human ground truth — labels are model output and may contain errors.

Two text fields per block:

• —`ocr` — the teacher's raw output, captured before any rendering post-processing.
• —`ocr_revised` — Qwen3.6 correction of the raw OCR against the crop image (populated for 99.99% of rows). The reviser runs with a corpus-specific system prompt that states the text is a 1980s–90s Taiwanese legal newspaper in vertical layout read right-to-left — this fixes systematic raw-OCR reading-order errors (e.g. a masthead read left-to-right as 「版一第」 is corrected to 「第一版」) — plus the issue number/date as reference metadata. Unlike the sibling TaiwanFinancial dataset there is no PDF text layer (pure scans), so revision is grounded in the image alone.

Both text fields are normalized to Traditional Chinese with OpenCC (s2t): the source is Traditional-only, so any Simplified characters the VLMs emitted are OCR errors and are converted back (46% of values had at least one such fix — vertical bilevel scans confuse scripts notably more than print PDFs).

Schema

Type distribution

text 80.5% · doc_title 11.6% · header 5.8% · image 1.0% · paragraph_title 0.9% · plus trace footer, table, number, seal, … (newspaper pages are text-dominant; tables/seals are rare in this corpus).

Content character

Legal-domain vocabulary: court announcements, case reports, judicial-personnel news, statute commentary, 司法院 official notices. Era-typical typography: letterpress with worn glyphs, dense multi-column vertical layout, calligraphic mastheads (the 1981 創刊號 included), name lists in fine print.

Usage

from datasets import load_dataset
ds = load_dataset("andynoodles/JudicialYuanPublication", split="train")
row = ds[0]
row["image"]                       # PIL crop
row["type"], row["ocr_revised"]    # prefer ocr_revised; fall back to ocr

Quality & limitations

• —Model-generated labels (distillation) — not human-verified, and grounded in the crop image only (no text layer exists for these scans).
• —The raw ocr field frequently mis-orders vertical text (reads left-to-right); ocr_revised corrects most of this but is not guaranteed. Prefer `ocr_revised`.
• —Bilevel JBIG2 scans: small print can be degraded; rare/variant legal characters may be mis-read by both models.
• —image-type rows (1.0%) are photographs/figures; their OCR fields carry the teacher's description output, not transcription.

Acknowledgements

• —[PP-DocLayoutV3](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3_safetensors) — document layout detection.
• —[PaddleOCR-VL-1.6](https://huggingface.co/PaddlePaddle/PaddleOCR-VL-1.6) — vision-language OCR (the ocr field).
• —[Qwen3.6-35B-A3B-FP8](https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8) — OCR revision (ocr_revised).
• —[OpenCC](https://github.com/BYVoid/OpenCC) — Simplified→Traditional conversion.

License & provenance

Source documents were downloaded from the 司法院 (Judicial Yuan) public website archive, whose open-data declaration adopts 政府資料開放授權條款第1版 (Open Government Data License v1.0, Taiwan) — free reuse including commercial, sublicensable, irrevocable, attribution required: 資料來源：司法院. Court and government documents are additionally excluded from copyright by Taiwan Copyright Act §9. Labels are generated by PaddleOCR-VL and Qwen3.6; verify fitness for your use case.

Citation (required)

This dataset is part of the Taiwan Document OCR (zh-Hant) collection by Andy Chen (andynoodles). If you use this dataset — in research, training data, products, or derivatives — you must cite it:

@misc{chen2026_judicial_weekly,
  title        = {Judicial Yuan Weekly OCR (block-level, zh-Hant, vertical text)},
  author       = {Chen, Andy (andynoodles)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/andynoodles/JudicialYuanPublication}},
  note         = {Part of the Taiwan Document OCR (zh-Hant) collection}
}

Citing this dataset is required in addition to the source-attribution noted under License & provenance above. Please retain both this citation and the original source attribution in any redistribution or derived work.