Datasets:
PMC-MAD: Missing-Abstract Distribution-aligned PMC
Dataset Description
PMC-MAD is a biomedical full-text dataset designed for the Biomedical Abstract Generation (BAG) task. It contains 46,309 articles drawn from the PubMed Central (PMC) Open Access subset, stratified to reflect the publication-type distribution of PubMed articles that lack abstracts. The dataset was constructed to support training and evaluation of abstract generation models targeting article types that usually lack abstracts.
- Paper: Divide-Prompt-Refine: a Training-Free, Structure-Aware Framework for Biomedical Abstract Generation (under review)
Dataset Summary
As of April 2026, approximately 29% of PubMed articles (11.6M out of 40.4M) lack abstracts. This gap is concentrated in publication types like case reports, editorials, and letters, which carry substantial scientific value but are excluded from most existing summarization datasets. PMC-MAD addresses this by sampling PMC OA articles in proportions that mirror the abstract-less PubMed distribution, providing a more representative benchmark for downstream BAG applications.
The dataset spans publications from 1987 to 2023.
Supported Tasks
- Abstract generation — generating abstracts from biomedical full-text articles, particularly for publication types where abstracts are commonly missing.
Languages
English (en).
Dataset Structure
Data Instances
Each instance corresponds to one PMC OA article and contains the article's PMCID, title, full text, reference abstract, and section-level structure.
Data Fields
| Field | Type | Description |
|---|---|---|
article_id |
string |
PMCID of the source article (e.g., PMC6625196). |
title |
string |
Article title. |
abstract |
string |
Reference abstract written by the original authors. |
full_text |
string |
Concatenated full text of the article body. |
sections |
List[List[str]] |
Article body organized as a list of sections, where each section is a list of paragraph strings. |
section_names |
List[str] |
Section headers aligned with sections (same length and order). |
Data Splits
| Split | Articles |
|---|---|
| train | 37,047 |
| validation | 4,631 |
| test | 4,631 |
| Total | 46,309 |
The split ratio is approximately 80/10/10.
Dataset Creation
Curation Rationale
Existing biomedical summarization datasets (e.g., the PubMed Summarization dataset) draw primarily from research articles with structured abstracts and do not reflect the distribution of articles that actually lack abstracts in PubMed. PMC-MAD is constructed via stratified sampling so that its publication-type distribution aligns with that of abstract-less PubMed articles, making it a more realistic benchmark for the BAG task.
Source Data
Articles were sampled from the PubMed Central Open Access subset. Publication-type queries were adapted from Menke et al. (2024) to estimate the publication-type distribution of abstract-less PubMed articles, and stratified sampling was used to retrieve 130,000 candidate XML files reflecting that distribution.
Preprocessing
- XML parsing followed an extraction pipeline adapted from the Longsummarization framework (Cohan et al., 2018).
- Articles that were unparseable or lacked extractable abstracts were filtered out, leaving 46,309 articles.
- HTML/XML markup was removed from text fields.
Annotations
The dataset contains no additional manual annotations beyond the original article structure (titles, abstracts, sections) extracted from PMC XML.
Considerations for Using the Data
Distribution Bias
The publication-type distribution is intentionally aligned with abstract-less PubMed articles, not with PubMed as a whole. Models trained or evaluated on PMC-MAD may therefore behave differently on general PubMed corpora. This is by design, but users targeting general-purpose biomedical summarization should be aware of this skew.
Document Length
Article lengths follow a heavy long-tail distribution: median length is around 2,959 tokens, but the longest article exceeds 1.18M tokens. Approximately 22% of articles exceed 8,192 tokens, which is the practical limit for many long-document models on standard hardware. Users may need to truncate or further filter depending on their model's context window.
Discourse Structure
The dataset is best suited to articles whose discourse roughly follows the BOMRC schema. Articles with non-standard structure (e.g., some editorials, letters) may be less well-served by structure-aware methods that assume this organization.
Citation
Will add citation here when the paper is published.
- Downloads last month
- 8