gr8monk3ys/academic-papers-dataset

CS/ML Academic Papers Dataset

A curated dataset of academic paper metadata sourced from arXiv, spanning key Computer Science and Machine Learning categories. Designed for research in text classification, topic modelling, summarisation, and scientometric analysis.

Dataset Description

Overview

The published snapshot (collected 2026-07-12) contains metadata for 2,019 papers — 2,500 were fetched and 481 cross-listed duplicates removed — across five core arXiv categories at the intersection of computer science and machine learning:

Category	Description
cs.AI	Artificial Intelligence
cs.CL	Computation and Language (NLP)
cs.CV	Computer Vision and Pattern Recog.
cs.LG	Machine Learning
stat.ML	Statistics — Machine Learning

Papers are sorted by submission date (most recent first) and deduplicated by arXiv ID, so cross-listed papers appear only once.

Collection Methodology

1. The arXiv API is queried via the arxiv Python package.
2. For each of the five categories, the 500 most recent papers are fetched (configurable).
3. Results are deduplicated on arxiv_id — papers cross-listed in multiple categories keep only their first occurrence.
4. Rate limiting (3.5 s between pages, 5 retries on failure) ensures compliance with arXiv's API Terms of Service.

The collection script is fully reproducible: see create_dataset.py.

Schema

Column	Type	Description
arxiv_id	string	arXiv identifier (e.g. 2401.12345v1)
title	string	Paper title, whitespace-normalised
abstract	string	Full abstract text, whitespace-normalised
authors	list[string]	Ordered list of author names
categories	list[string]	All arXiv categories assigned to the paper
primary_category	string	The primary arXiv category
published	string	ISO-8601 publication timestamp
updated	string	ISO-8601 last-updated timestamp
doi	string	Digital Object Identifier (empty string if unavailable)
url	string	Canonical arXiv abstract URL

Splits

Split	Rows	Share
train	1,817	90 %
test	202	10 %
Total	2,019

The split is performed with a fixed random seed (42) for reproducibility. Row counts describe the 2026-07-12 snapshot; re-running the collection script on a different date will produce a different corpus (see Limitations).

Intended Uses

• Text classification — predict the primary arXiv category from the title or abstract.
• Summarisation — generate concise summaries from abstracts or use titles as abstractive targets.
• Topic modelling — discover latent research themes across categories.
• Scientometrics — analyse publication trends, author collaboration patterns, and category overlap.
• Information retrieval — build search or recommendation engines over academic literature.

Example

from datasets import load_dataset

ds = load_dataset("gr8monk3ys/cs-ml-academic-papers")

# Peek at the first training example
print(ds["train"][0])
# {
#     "arxiv_id": "2401.12345v1",
#     "title": "An Example Paper on Large Language Models",
#     "abstract": "We propose a novel method for ...",
#     "authors": ["Alice Researcher", "Bob Scientist"],
#     "categories": ["cs.CL", "cs.AI"],
#     "primary_category": "cs.CL",
#     "published": "2024-01-20T18:00:00+00:00",
#     "updated": "2024-01-22T12:30:00+00:00",
#     "doi": "10.1234/example.2024.001",
#     "url": "http://arxiv.org/abs/2401.12345v1"
#  }

Reproducing the Dataset

# Install dependencies
pip install -r requirements.txt

# Fetch papers and save locally (Parquet + HF Dataset format)
python create_dataset.py

# Optionally push to the HuggingFace Hub
python create_dataset.py --push --hf-repo gr8monk3ys/cs-ml-academic-papers

# Run exploratory data analysis and generate plots
python explore_dataset.py

CLI Options — create_dataset.py

Flag	Default	Description
--per-category	500	Papers to fetch per arXiv category
--push	off	Push dataset to HuggingFace Hub
--hf-repo	gr8monk3ys/cs-ml-academic-papers	Hub repository ID
--output-dir	./data	Local output directory
--verbose	off	Enable debug logging

CLI Options — explore_dataset.py

Flag	Default	Description
--data-dir	./data	Local data directory
--from-hub	—	Load from HF Hub repo instead
--plots-dir	./plots	Output directory for PNG plots
--verbose	off	Enable debug logging

Generated Visualisations

The EDA script (explore_dataset.py) produces the following plots:

• category_distribution.png — horizontal bar chart of paper counts per primary category.
• abstract_length_histogram.png — word-count distribution of abstracts with mean/median markers.
• abstract_length_by_category.png — box plots comparing abstract lengths across categories.
• authors_per_paper.png — histogram of author counts.
• publication_timeline.png — monthly publication volume over time.
• tfidf_terms_heatmap.png — heatmap of the top TF-IDF terms per category.

Limitations and Bias

• The dataset is a snapshot — arXiv receives thousands of new submissions daily; re-running the collection script at different times will produce different results.
• Only five categories are included. Many relevant papers (e.g., cs.IR, cs.RO, math.OC) are excluded.
• Papers cross-listed under multiple categories are kept only once, so secondary categories are under-represented in the primary-category distribution.
• Metadata only — full paper text and PDFs are not included.

Citation

If you use this dataset in your work, please cite it as:

@misc{scaturchio2024csml_papers,
  author       = {Lorenzo Scaturchio},
  title        = {{CS/ML Academic Papers Dataset}},
  year         = {2024},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/gr8monk3ys/cs-ml-academic-papers}},
}

License

This dataset is released under the MIT License. The underlying paper metadata is sourced from arXiv and is subject to arXiv's Terms of Use.