---
task_categories:
  - text-classification
language:
  - en
tags:
  - dual-use
  - gradient-routing
  - modular-pretraining
size_categories:
  - 100K<n<1M
---

# Dual-Use Papers (metadata only)

A curated collection of academic papers labeled by subject-matter category.
Built for research on dual-use capabilities in LLMs.

## Schema

| Column     | Type   | Description                                    |
|------------|--------|------------------------------------------------|
| `category` | string | Topic label (see mapping below)                |
| `source`   | string | Provenance tag, e.g. `europe-pmc`, `arxiv_cs.CR`, `osti` |
| `url`      | string | URL to the original PDF                        |

## Category Mapping

| Source                         | Category   |
|--------------------------------|------------|
| Europe PMC                     | `biology`  |
| OSTI                           | `nuclear`  |
| arXiv `nucl-th`, `nucl-ex`     | `nuclear`  |
| arXiv `cs.CR`, `cs.NI`, `cs.OS`| `cyber`    |
| arXiv (all other categories)   | `other`    |

## Statistics

**Total rows: 415,119**

| Category   |   Rows |
|------------|-------:|
| `biology`  | 192,403 |
| `cyber`    |  49,806 |
| `nuclear`  |  79,351 |
| `other`    |  93,559 |

| Source        |   Rows |
|---------------|-------:|
| `europe-pmc`  | 192,403 |
| `osti`        |  31,963 |
| arXiv (cyber) |  49,806 |
| arXiv (nuclear) | 47,388 |
| arXiv (other) |  93,559 |

## Sources

- **Europe PMC** — Open-access biomedical and life-science literature from
  [europepmc.org](https://europepmc.org).
- **arXiv** — Preprints across physics, mathematics, computer science, and more
  from [arxiv.org](https://arxiv.org). Dual-use categories (cyber, nuclear) are
  collected exhaustively from 1991 to the present. Other categories are sampled
  and balanced to match the dual-use count.
- **OSTI** — Technical reports and journal articles from the U.S. Department of
  Energy's [Office of Scientific and Technical Information](https://www.osti.gov).

## Processing

1. Metadata is collected via source-specific APIs (arXiv, Europe PMC, OSTI).
2. arXiv "other" papers are shuffled and truncated so their count exactly matches
   the number of dual-use arXiv papers.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("AE-data/dual-use-papers", split="train")
print(ds)
print(ds[0]["category"], ds[0]["url"])
```

## Note

This dataset is intended solely for safety and scientific research purposes.