ttkhiem/introds-m2

introDS Group 15 - arXiv dataset, range of m2

Slice of the Introduction to Data Science term project (HCMUS) of Group 15, built with introds. Report: https://ttphat-fitus.github.io/IntroDS-Book/.

It covers the arXiv papers 2608.29007 to 2609.02833 (5,000 papers), every version of each paper, and their references from Semantic Scholar.

Table	Rows	Content	Milestone
papers	5,000	Crawled paper metadata	1
references	82,021	Crawled references that have an arXiv ID, one row per (citing paper, reference)	1
elements	2,384,659	Deduplicated parsed elements across versions	2
hierarchy	2,795,419	Per-version node-parent edges	2
bibtex	197,592	Unified BibTeX entries after deduplication	2
predictions	-	Reference matching results (top-5 candidates)	3 (to come)

Schemas

papers: arxiv_id (string, non-null), title (string), authors (list of strings), submission_date (ISO 8601 string, date of v1), revised_dates (list of strings, dates of v2..vn), venue (string, nullable; arXiv journal-ref, else Semantic Scholar venue).

references: arxiv_id (string, non-null, citing paper), ref_arxiv_id (string, non-null), ref_title (string), ref_authors (list of strings), ref_submission_date (string), ref_s2_id (string, Semantic Scholar paper ID). Only references that have an arXiv ID.

elements: element_id (string, non-null, <yymm-nnnnn>#<prefix><n>), arxiv_id (string, non-null), kind (dictionary: document, heading, list, item, sentence, formula, figure), content (string; cleaned LaTeX, inline math as $...$, block math as \begin{equation}...). Elements whose cleaned content is identical across versions share one ID.

hierarchy: arxiv_id (string, non-null), version (int), element_id (string, non-null), parent_id (string, non-null). Each version has its own root (#root1, #root2, ...), whose parent is itself.

bibtex: arxiv_id, citation_key (non-null), entry_type (dictionary), title, authors (list of strings), year (int, nullable), doi, journal, url. One row per reference after deduplication across versions; \cite{} keys in elements use these keys.

Usage

from datasets import load_dataset

papers = load_dataset("ttkhiem/introds-m2", "papers", split="train")
elements = load_dataset("ttkhiem/introds-m2", "elements", split="train", streaming=True)