langminer/doc-content-clustering-740

Document Content-Clustering Benchmark (12 classes, 740 items)

A small, curated benchmark for clustering documents by their content topic (not by their visual form/layout). Each item is a single document page provided as an image plus two text views (a VLM description and OCR markdown), with a ground-truth content class.

The set is intentionally "tangle-stripped": 27 borderline items whose content sits ambiguously between two classes were removed from a larger 907-item pool to leave 740 cleanly-labelled items across 12 classes.

Classes (740 items)

label	class	n
0	academic-bio	19
1	academic-crosstopic	18
2	academic-math	63
3	academic-medicine	45
4	academic-physics	44
5	academic-psych	40
6	balance-sheet	80
7	corporate-filing	100
8	diagram-schematic	108
9	payslip	83
10	resume	80
11	statement	60

Fields

field	type	description
image	image	the document page (PNG/JPG)
label	ClassLabel	integer class id (see table)
label_name	string	class name, redundant with label
stem	string	stable per-item source id / filename stem
desc	string	a vision-language-model description of the page, wrapped in [DESC]…[DESC]
md	string	OCR text of the page as markdown, wrapped in [MARKDOWN]…[MARKDOWN] (≤6000 chars)
desc_md	string	desc + md concatenated (the "shipped" text view used for clustering)

Usage

from datasets import load_dataset

ds = load_dataset("langminer/doc-content-clustering-740", split="train")
print(ds)
print(ds.features["label"].names)

ex = ds[0]
ex["image"]        # PIL image
ex["label_name"]   # e.g. "academic-medicine"
ex["desc_md"]      # text view used for content clustering

Sources & provenance

Assembled from public / openly-listed sources; each class was fetched programmatically:

• academic-* — arXiv (per-paper license, research use). The six academic-* classes are a fine-grained split of arXiv first pages by discipline (biology, cross-topic, math, medicine, physics, psychology).
• balance-sheet, corporate-filing — SEC EDGAR (US-gov public filings, public domain).
• diagram-schematic — Wikimedia Commons (per-file free license).
• resume — mirrored from the public HF dataset opensporks/resumes (Kaggle snehaanbhawal/resume-dataset, LiveCareer resume snapshots).
• statement — pages rendered from the public HF dataset Akashved/Indian-Bank-Statements.
• payslip — payslip pages.

Licensing follows the underlying sources on a per-item basis; treat the release as research use. If you are a rights-holder for any item and want it removed, open an issue on the repo.

Note on personal data: the resume, statement, and payslip classes are document types that can carry personal information. Use for research only; do not attempt to re-identify or contact any individual.