Datasets:
File size: 2,410 Bytes
b2ecf4b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | ---
license: mit
task_categories:
- text-classification
language:
- en
- multilingual
tags:
- abstract-detection
- scientific-text
- quality-filtering
- pubverse
size_categories:
- 1K<n<10K
---
# Abstract Archon Training Data
Binary classification dataset for detecting whether a text is a real scientific research abstract or non-abstract content (figure captions, supplementary material references, author bylines, journal metadata, HTML artifacts, taxonomy stubs).
## Dataset Details
- **Positive examples (label=1)**: 2,000 real abstracts randomly sampled from a 198M publication database (publications not flagged by any quality filter, length >= 200 characters)
- **Negative examples (label=0)**: ~2,000 non-abstract texts from various garbage categories:
- `figure_table_caption`: Figure and table captions misidentified as abstracts
- `journal_article_scrape`: Scraped article metadata (titles, access info, author lists)
- `html_heavy_text`: Content with substantial HTML markup
- `author_byline`: Author affiliations and bylines
- `moesm_title`: Electronic supplementary material descriptions
- `supplementary_content`: Supplementary material references
- `taxonomy_stub`: Taxonomic database stubs
## Format
NDJSON with fields:
- `text`: First 500 characters of the abstract field
- `label`: 1 (real abstract) or 0 (garbage/non-abstract)
- `source`: Category label (e.g., `positive_real_abstract`, `negative_figure_table_caption`)
- `source_id`: OpenAlex-style source identifier
## Data Curation
Negatives were manually curated to remove misclassified entries. The `html_heavy` category was entirely removed after review showed nearly all entries were real abstracts with minor HTML entity artifacts. The `supplementary_content` category was filtered to retain only entries that begin with supplementary material indicators (e.g., "Figure S1", "Additional file", "Supplementary dataset").
## Intended Use
Training a lightweight binary classifier (Potion-32M embeddings + LogisticRegression) as a quality gate for large-scale scientific publication databases. The classifier identifies non-abstract text that has been incorrectly stored in abstract fields.
## Source
Sampled from a PostgreSQL database of ~198M publications aggregated from OpenAlex, PubMed, arXiv, bioRxiv, and medRxiv. Part of the [PubVerse](https://github.com/jimnoneill/pubverse) project.
|