taesiri/ArXivSignals-FullText

ArXivSignals FullText — arXiv Papers OCR'd to Markdown + Layout

A continuously-updated, day-partitioned dataset of arXiv papers converted to clean full text by a vision OCR pipeline: each paper's PDF is rendered to Markdown (headings, paragraphs, tables as HTML, math as LaTeX) plus a structured layout JSON (typed, bounding-boxed blocks). It is the full-text companion to taesiri/ArXivSignals (metadata + LLM signal & summaries) and joins it on paper_id.

How it's made

• Source: arXiv PDFs (publicly available).
• OCR: a document vision-language model (Baidu-family "Unlimited-OCR" via MLX), run at 4-bit precision, page-by-page, producing Markdown + a raw tagged form + a typed layout tree. No LLM rewriting — this is transcription, not summarization.
• Scope: papers that are also in the public ArXivSignals catalog. The corpus fills in continuously (newest-first); coverage grows daily.

Schema (papers config, partitioned by announce_date)

Column	Type	Notes
paper_id	string	arXiv id (joins taesiri/ArXivSignals)
announce_date	date	arXiv announcement date (partition key)
title, abstract	string	arXiv metadata
author_names	list	display names
authors_json	string	full author structure (JSON)
categories	list	arXiv categories
primary_category	string
ocr_markdown	string	the OCR'd full text (Markdown; HTML tables; LaTeX math)
ocr_layout	string	typed + bbox'd layout blocks (JSON: {items:[{page,type,bbox,content}]})
pages	int	page count
md_chars, n_layout_blocks	int	content stats
ocr_tokens, ocr_tps, ocr_duration_s	numeric	OCR throughput stats
ocr_precision, ocr_model	string	OCR configuration
ocr_finished_at	string	when this paper was OCR'd

from datasets import load_dataset
ds = load_dataset("taesiri/ArXivSignals-FullText", "papers", split="corpus")
print(ds[0]["ocr_markdown"][:500])

Licensing, attribution & takedown

This is important — read before redistributing. The ocr_markdown / ocr_layout fields are a machine-generated transcription of arXiv PDFs. arXiv papers are licensed individually by their authors (arXiv's default non-exclusive license, or CC-BY / CC-BY-SA / CC0 / other, per submission), and that license governs the underlying content of the OCR text. This dataset does not grant any rights beyond those of each source paper.

• Metadata (title, abstract, authors, categories) is provided under CC-BY-4.0, consistent with the companion catalog dataset.
• OCR full text is provided for research and text/data-mining purposes, as a derived representation of publicly available papers. Redistribution or reuse of any paper's text is subject to that paper's own license — check it before reusing.
• Attribution: always cite the original arXiv paper (paper_id), not this dataset, as the source of the content.
• OCR is imperfect: expect errors in math, tables, multi-column, and scanned pages. Treat the text as machine-transcribed, not authoritative.
• Takedown / opt-out: if you are an author (or rights holder) and want a paper removed, open an issue / discussion on this dataset repo — it will be removed promptly.

Maintained by @taesiri · powers arxivsignals.io.