yuiseki/gsj-reports

GSJ Bulletin text: 地質調査所月報 and 地質調査研究報告 (1950-2026)

Page-level text of two journals of the Geological Survey of Japan (GSJ), AIST, extracted from the text layer of the official PDFs. One JSON line is one PDF page.

This dataset contains only text. The PDFs, figures, maps and photographs are not included; every line gives the official PDF URL (source_url) and the SHA-256 of the PDF that was read (source_pdf_sha256), so the original can be downloaded and checked.

This is not an official GSJ product and is not endorsed by GSJ.

The two series

config	file	journal	years	volumes	PDFs	pages	characters	empty pages	size
monthly	monthly.jsonl	地質調査所月報 (Bulletin of the Geological Survey of Japan)	1950-2001	Vol. 1 No. 1 to Vol. 52 No. 2-3 (595 issues)	3,514	38,366	212,624,609	107	313,487,344 bytes
report	report.jsonl	地質調査研究報告 (Bulletin of the Geological Survey of Japan)	2001-2026	Vol. 52 No. 4 to Vol. 77 No. 2 (147 issues)	652	10,699	49,895,422	961	72,569,746 bytes
default	both files	both	1950-2026		4,166	49,065	262,520,031	1,068	386,057,090 bytes

地質調査研究報告 is the successor of 地質調査所月報 after the Geological Survey of Japan became part of AIST in 2001; the volume numbering continues (月報 ends at Vol. 52 No. 2-3, 研究報告 starts at Vol. 52 No. 4). The series field (monthly / report) only records which index the PDF was listed in. Load one series with load_dataset("yuiseki/gsj-reports", "monthly") or "report"; the default config reads both files.

Coverage is every individual PDF linked from the two official indexes on 2026-10-06:

• 地質調査所月報: https://www.gsj.jp/publications/pub/bull-gsj/index.html
• 地質調査研究報告: https://www.gsj.jp/publications/bulletin/index.html

This includes articles, covers, tables of contents, errata and separately published figure/table PDFs. Whole-issue _full.pdf files were skipped because they duplicate the individual PDFs.

Most text is Japanese. English titles, abstracts and figure captions are common, and some articles are entirely in English: by a simple character count about 3,593 pages of monthly and 3,174 pages of report are mostly Latin script.

Files and order

• monthly.jsonl, report.jsonl: UTF-8 JSON Lines, one page per line, separated by \n only. Documents are in the order of the source catalog (volume and issue ascending, PDFs in the order the issue lists them), pages in PDF order. Each line is the per-PDF extraction line copied byte for byte; nothing was re-serialised. Split on \n, not on every Unicode line break, if you read the files yourself.
• MANIFEST.json: lines, documents, bytes and SHA-256 of each file; pages, characters, empty pages and U+FFFD counts of each series.
• provenance.yaml: source indexes, catalog date, fetch and extraction dates, tool versions, and what was changed (nothing in the text).
• LICENSE: licence, quotations of the GSJ licence pages, attribution.

Hugging Face converts the JSONL to Parquet automatically for the viewer and datasets; no Parquet is published separately.

Schema

All 17 fields are present on every line.

field	type	meaning
document_id	string	SHA-256 (hex) of source_url. Same for every page of one PDF.
series	string	monthly (地質調査所月報) or report (地質調査研究報告).
source_url	string	Official URL of the PDF on www.gsj.jp.
issue_url	string	Official issue page that links the PDF.
source_pdf_path	string	Path of the PDF in the compiler's local collection: pdf/<series>/ plus the last part of source_url, percent-decoded (e.g. pdf/monthly/18-02_02-Tab.1(1).pdf for .../18-02_02-Tab.1%281%29.pdf). The PDFs are not in this dataset; the field only names the file under the official URL's name.
source_pdf_sha256	string	SHA-256 of the PDF that was downloaded and read.
source_pdf_bytes	int	Size of that PDF in bytes.
page_number	int	1-based physical page index in the PDF. Not the printed page number.
text	string	Text of the page from pdftotext -layout, as extracted. Empty string for a page without a text layer.
characters	int	Length of text in Unicode code points.
non_whitespace_characters	int	Code points of text that are not whitespace.
has_text	bool	non_whitespace_characters > 0.
needs_ocr_review	bool	not has_text. Marks empty pages only; it does not detect bad text layers.
replacement_characters	int	Number of U+FFFD in text.
extraction_method	string	Always pdftotext-layout.
extractor_version	string	pdftotext version 24.02.0 (4,158 PDFs) or pdftotext version 24.08.0 (8 PDFs).
rights_review	string	Constant note in Japanese written at extraction time: GSJ's general terms are CC BY; third-party material and quoted text need individual checks.

How the text was made

1. The PDF URLs were collected from the two indexes (catalog of 2026-10-06, 4,166 URLs) and each PDF was downloaded once, sequentially, recording its size and SHA-256 (2026-10-06).
2. Each PDF was converted with Poppler pdftotext -enc UTF-8 -layout, which reads the PDF's own text layer. The output was split on form feeds into pages and the page count was checked against pdfinfo. No new OCR was run, and no correction, normalisation (width, old kanji, punctuation), header removal or de-hyphenation was applied. Whitespace from the layout is kept.
3. The per-PDF JSONL files were concatenated into one file per series. All 4,166 PDFs were extracted without errors.

Quality caveats

The text is the PDF text layer as it is. It is not cleaned training text.

• Older volumes appear to be scans with an OCR text layer, and it contains misrecognised characters. Examples: the first article of Vol. 1 No. 1 (1950) is mostly symbol noise on its figure pages; a 1971 page reads にっいて for について and 銅ンリチウム for 銅，リチウム. Absence of U+FFFD and has_text = true do not mean the text is correct.
• -layout keeps the physical layout. On two-column pages the left and right columns are on the same text line, separated by runs of spaces, so a line of text alternates between columns. Split on wide gaps if you need continuous reading order.
• Running heads (journal name, volume, article title, authors), page numbers, captions, table cells and text inside figures are mixed into the page text.
• 1,068 pages are empty (has_text = false, needs_ocr_review = true): 107 in monthly and 961 in report. 33 report PDFs (831 pages) and 90 single-page monthly PDFs have no text layer at all. Their content is only in the PDF.
• report has 55,906 U+FFFD characters in 27 PDFs, where a font has no usable Unicode mapping (for example tables in Vol. 59 No. 9).
• A PDF is not always an article: it can be a cover, a contents list, an erratum or a figure or table sheet.

Getting and checking the original PDF

source_url is the official PDF. Download it and compare its SHA-256 with source_pdf_sha256 (and its size with source_pdf_bytes). If they differ, GSJ has replaced the file since 2026-10-06 and the text may not match. page_number is the page index in that PDF.

curl -sO "https://www.gsj.jp/data/bull-gsj/01-01_01.pdf"
sha256sum 01-01_01.pdf   # b7fa9561b97e2cd21e907a5e42308fb0cdf9a2ffa60a64d93d0239b9204e1473

Licence

CC BY 4.0. The source journals are published by the Geological Survey of Japan (GSJ), AIST, and the licence comes from GSJ's own terms (read on 2026-10-07; full quotations in LICENSE):

• GSJ licence table https://www.gsj.jp/license/docs/gsj_pubs_lcs.pdf lists both 地質調査研究報告 and 地質調査所月報 under 「地球科学成果報告書 （ライセンス CC BY *1）」. Its footnote *1: 「他機関の著作物が含まれる場合（例えば国土地理院の地形図やグーグルマップを基図としている場合など）には、別途、その機関からの許諾が必要になることがあります。」 (where works of other organisations are included, such as GSI topographic maps or Google Maps used as base maps, separate permission from that organisation may be needed).
• GSJ site rules https://www.gsj.jp/license/license.html: printed publications are in principle under Creative Commons; the rules are compatible with CC BY 4.0 and content may be used under CC BY (1. 7) ウ). Users must state the source (1. 1) ア), must state any editing or processing and must not present edited content as made by GSJ (1. 1) イ). Third parties may hold rights in some content, which is not always marked; obtaining permission and checking is the user's responsibility (1. 2) ア/イ).

This dataset contains only the extracted text, not the figures, maps or PDFs. Text inside the articles that is quoted from, or owned by, third parties keeps its own rights and is not licensed by GSJ or by this dataset.

Attribution

In the form requested by the GSJ site rules (1. 1) ア/イ):

出典：地質調査所月報（産総研地質調査総合センター）（https://www.gsj.jp/publications/pub/bull-gsj/index.html） 出典：地質調査研究報告（産総研地質調査総合センター）（https://www.gsj.jp/publications/bulletin/index.html） 地質調査所月報・地質調査研究報告（産総研地質調査総合センター）の各PDF（各行の source_url）の文字層から pdftotext で本文を抽出し、yuiseki がページ単位の JSONL に収録したものである（Hugging Face: yuiseki/gsj-reports）。本文の訂正・正規化は行っていない。

In English: Source: 地質調査所月報 and 地質調査研究報告, Geological Survey of Japan, AIST. Text extracted with pdftotext from each PDF's text layer and compiled into page-level JSONL by yuiseki (yuiseki/gsj-reports); not corrected or normalised; not made or endorsed by GSJ.

When you cite a single article, cite the article (authors, title, journal, volume, issue, pages) in the style of your venue and give its source_url.