README.md

metadata

license: cc0-1.0
task_categories:
  - text-classification
task_ids:
  - multi-label-classification
tags:
  - extreme-multi-label
  - pubmed
  - mesh
  - biomedical
  - nlp
language:
  - en
pretty_name: PubMed MultiLabel Text Classification (MeSH)
size_categories:
  - 10K<n<100K

PubMed MultiLabel Text Classification (MeSH)

A dataset of 50,000 PubMed biomedical articles, each manually annotated by domain experts with MeSH (Medical Subject Headings) labels. With 21,918 unique labels and a mean of ~12.7 labels per document, this is a densely-labeled extreme multi-label classification benchmark.

Dataset Description

Property	Value
Train examples	40,000
Test examples	10,000
Total unique MeSH labels	21,918
Mean labels per document	~12.7
Median labels per document	12
Max labels per document	46

Label distribution

Docs per label	# Labels	% of total
1	3,990	18.2%
2–5	7,020	32.0%
6–10	3,412	15.6%
11–50	5,518	25.2%
51–100	1,068	4.9%
101+	910	4.2%

MeSH Root Categories

Each label belongs to one or more MeSH root categories. The dataset includes binary indicator columns for the 14 root categories:

Code	Root Category
A	Anatomy
B	Organisms
C	Diseases
D	Chemicals and Drugs
E	Analytical, Diagnostic and Therapeutic Techniques, and Equipment
F	Psychiatry and Psychology
G	Phenomena and Processes
H	Disciplines and Occupations
I	Anthropology, Education, Sociology, and Social Phenomena
J	Technology, Industry, and Agriculture
L	Information Science
M	Named Groups
N	Health Care
Z	Geographicals

Fields

Field	Type	Description
pmid	string	PubMed article ID
title	string	Article title
abstract	string	Article abstract text
label_ids	list[int]	MeSH label indices (into the 21,918-label vocabulary)
label_names	list[string]	Human-readable MeSH label names
mesh_roots	dict	Binary flags {"A": 0/1, ..., "Z": 0/1} for root categories

Additional files

• label_vocab.json — ordered list of all 21,918 MeSH label names (index = label ID)
• label_metadata.jsonl — full label metadata including MeSH tree IDs and root categories for hierarchical classification research

Splits

An 80/20 random split with seed 42 (no predefined split exists in the original data).

Usage

from datasets import load_dataset

ds = load_dataset("Tellurio/PubMed-MultiLabel-MeSH")

example = ds["train"][0]
print(example["title"])
print(example["label_names"])  # e.g. ["Humans", "Female", "DNA Probes, HPV", ...]
print(example["label_ids"])    # e.g. [5, 2, 0, ...]
print(example["mesh_roots"])   # e.g. {"A": 0, "B": 1, "C": 1, ...}

Loading label metadata for hierarchical / zero-shot approaches

Each of the 21,918 MeSH labels has associated tree IDs and root categories stored in label_metadata.jsonl.

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Tellurio/PubMed-MultiLabel-MeSH",
    filename="label_metadata.jsonl",
    repo_type="dataset",
)

labels = []
with open(path) as f:
    for line in f:
        labels.append(json.loads(line))

# Example label entry
print(labels[0])
# {"id": 0, "label": "DNA Probes, HPV", "mesh_tree_ids": ["D13.444...", ...], "mesh_roots": ["Chemicals and Drugs [D]"]}

Source

Originally from Kaggle: PubMed MultiLabel Text Classification Dataset MeSH by Owais Ahmad.

Citation

@misc{pubmed_multilabel_mesh,
  author = {Owais Ahmad},
  title = {PubMed MultiLabel Text Classification Dataset MeSH},
  year = {2022},
  publisher = {Kaggle},
  url = {https://www.kaggle.com/datasets/owaiskhan9654/pubmed-multilabel-text-classification}
}

License

CC0: Public Domain