SIY Longitudinal Health-Statistics Extraction Framework: Database, Pipeline and Discrepancy Catalogue

Open dataset and Python pipeline accompanying a longitudinal table-extraction study on the seventeen-volume Turkish Ministry of Health Statistics Yearbook (Sağlık İstatistikleri Yıllığı, SIY, 2008–2024).This deposit provides: • a 69 MB unified SQLite database integrating four data layers — extraction output, curated indicator catalogue, inter-edition data-quality discrepancies, and a full-text page corpus; • a 24-script Python pipeline that reproduces every step of the cascade (rasterisation → Table-Transformer detection → structure recovery → LayoutLMv3 multimodal cell understanding → ByT5 byte-level encoding repair → schema normalisation); • all six figures of the accompanying manuscript and a graphical abstract; • a manuscript preprint (.docx).Released ground truth: 2,651 tables, 268,109 cells, across 17 editions.Curated layer: 46 indicators with 2,186 long-format observations.Data-quality layer: 63 documented inter-edition value discrepancies (34 major + 29 minor).Full-text page corpus: 7,725 pages (4,083 TR + 3,642 EN). Evaluation on the held-out 2024 test split (166 tables, 19,980 cells): TEDS-Struct 0.91, TEDS-Content 0.90, Cell-F1 0.87 — a 22-fold Cell-F1 improvement over pdfplumber baseline.License: data and figures under CC BY-SA 4.0; Python code under MIT. This deposit accompanies the manuscript "A Longitudinal Table Extraction Framework for Evidence-Based Health Policy" (PeerJ Computer Science, AI Application track, Manuscript ID CS-2026:05:141119). CONTENTS1. data/SIY_Veritabani_birlesik.db — Consolidated SQLite database (69 MB) containing the seven relations of the SIY extraction framework: yearbooks (17 editions), tables (2,651 tables), cells (268,109 ground-truth cells), indicators (46 curated indicators), values_long (2,186 long-format observations), discrepancies (63 inter-edition discrepancies), and full_text (7,725-page text corpus).2. code/ — Twenty-five Python scripts covering corpus acquisition (00_download_siy_pdfs.py), extraction (process_year.py / process_year2.py / build_summary_xlsx.py and supporting modules), the eighteen-script curation pipeline that produces the indicator catalogue and the discrepancy relation, and the standalone Docling baseline evaluator (code/evaluation/run_docling_baseline.py) added in v1.2.0.3. figures/Fig1_Pipeline.png — Architectural diagram of the cascade pipeline.4. README.docx — Step-by-step reproduction guide, including the modern document-AI baseline comparison protocol.5. CITATION.docx — Recommended citation and BibTeX entry.6. LICENSE-code.docx — MIT License covering all Python code.7. LICENSE-data.docx — Creative Commons Attribution-ShareAlike 4.0 International License covering the SQLite database and derived data. VERSION HISTORY- v1.2.0 (17 June 2026): Added run_docling_baseline.py and revised README to document the modern document-AI baseline comparison requested by reviewers.- v1.1.0 (16 June 2026): Initial public release accompanying the PeerJ submission. REPRODUCIBILITYThe 17-yearbook source PDF corpus (1.2 GB) is publicly redistributed by the Republic of Türkiye Ministry of Health and can be re-downloaded via code/00_download_siy_pdfs.py. The deposit is self-contained: a downstream analyst can re-run the full pipeline on the released PDFs, verify the 2,651 tables / 268,109 cells, reproduce the 63-entry discrepancy catalogue, and independently benchmark Docling on the same evaluation slice.