Datasets:

Tiansve
/

arxiv-cs-subfield

language:
  - en
license: cc-by-4.0
task_categories:
  - text-classification
size_categories:
  - 1K<n<10K
tags:
  - arxiv
  - computer-science
  - abstracts
  - multi-class

arXiv CS Sub-field Classification Dataset

Balanced, single-label classification dataset of recent arXiv abstracts across six overlapping CS sub-fields. Each row carries one paper's title and abstract, plus a label derived from its arXiv primary_category.

Supported tasks

text-classification — predict the arXiv sub-field given title + abstract.

Languages

English (en).

Data fields

field	type	description
`id`	string	arXiv identifier (e.g. `2401.12345v1`)
`title`	string	Paper title, whitespace-normalised
`abstract`	string	Paper abstract, whitespace-normalised, LaTeX preserved
`text`	string	`title + ". " + abstract`, the field used as model input
`label`	string	Human-readable sub-field name
`primary_category`	string	Original arXiv primary category
`published`	string	ISO timestamp of first arXiv submission

Label space

The six target arXiv categories (cs.CL, cs.LG, cs.CV, cs.IR, cs.AI, stat.ML) map to human-readable names:

arXiv primary_category	label
`cs.CL`	`cl_nlp`
`cs.LG`	`lg_ml`
`cs.CV`	`cv_vision`
`cs.IR`	`ir_retrieval`
`cs.AI`	`ai_general`
`stat.ML`	`stat_ml`

Splits

split	rows
train	3292
validation	412
test	412
total	4116

Splits are stratified 80/10/10 by label, seeded with random_state=42.

Source

arXiv API, queried with the arxiv Python package (Client(delay_seconds=3.5)).
Time range: 2023-01-01 to 2025-12-31.
Filtering rule: only papers whose primary_category equals the queried category are kept (so the label is unambiguous), then each class is down-sampled to the size of the smallest class for balance.

Annotation process

Labels are arXiv primary_category values; no manual annotation was performed. Boundary cases between e.g. cs.AI and cs.LG are inherently noisy and are part of what makes the task interesting.

Limitations

Single-label simplification of a fundamentally multi-label problem (most papers carry several arXiv categories).
Class boundaries are noisy by construction.
No temporal generalisation split — train, val and test all come from the same time window.

Licensing

arXiv abstracts are distributed under arXiv's non-exclusive license. This dataset redistributes metadata + abstracts for non-commercial research use. Each row keeps the arXiv id so users can verify the source.

Citation

@misc{arxiv-cs-subfield,
  author = { Tiansve },
  title  = { Tiansve/arxiv-cs-subfield },
  year   = 2026,
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Tiansve/arxiv-cs-subfield}}
}