Abu-Sameer-66/SciPeerBench

---
license: cc-by-4.0
task_categories:
- text-classification
language:
- en
tags:
- scientific-integrity
- fraud-detection
- peer-review
- research-ethics
- benchmark
size_categories:
- n<1K
---

# SciPeerBench v1.1

**World's first multi-dimensional scientific fraud detection benchmark.**

No other dataset labels papers across 14 fraud dimensions simultaneously.

## Stats

| Property | Value |
|----------|-------|
| Total papers | 644 |
| Fraud papers | 286 |
| Clean papers | 358 |
| Columns | 35 |
| Fraud dimensions | 14 |
| Year range | 1998–2026 |

## Why This Dataset is Unique

Every existing fraud dataset does binary labeling — fraud or not fraud.

SciPeerBench labels each paper across **14 dimensions** — statistical fraud,
figure manipulation, citation rings, LLM detection, and more.
Nobody has done this before.

## Categories

| Category | Count | Description |
|----------|-------|-------------|
| CONFIRMED_FRAUD | 286 | PubMed retracted + CrossRef verified |
| CLEAN | 223 | High quality multi-field papers |
| SUSPECTED_FRAUD | 109 | Journal expressions of concern |
| BORDERLINE | 16 | Famous disputed cases |
| BASELINE_ELITE | 10 | Nobel Prize and landmark papers |

## 14 Fraud Dimensions Per Paper

| Column | What it detects |
|--------|----------------|
| stat_audit_score | p-hacking, sample size issues |
| figure_forensics_score | image duplication, manipulation |
| methodology_score | causation claims, missing controls |
| citation_score | self-citation rings |
| reproducibility_score | code and data availability |
| novelty_score | incremental vs novel work |
| grim_score | mathematically impossible means |
| sprite_score | impossible distributions |
| granularity_score | Benford law violations |
| pcurve_score | publication bias |
| effect_size_score | inflated effect sizes |
| retraction_score | cited retracted papers |
| cartel_score | citation ring networks |
| llm_score | AI-generated paper detection |

## Famous Fraud Cases Included

- **Wakefield 1998** (The Lancet) — vaccines-autism data fabrication
- **LaCour 2014** (Science) — survey data never existed
- **Obokata 2014** (Nature) — STAP cells image manipulation
- **Stapel** — 58 psychology papers fabricated over 10 years

## Data Sources

- PubMed `Retracted Publication[pt]` — US National Library of Medicine
- CrossRef `update-type:retraction` — publisher verified DOI registry
- PubMed `Expression of Concern[pt]` — journal flagged papers
- Manual curation — court and media verified fraud cases

## Associated Project

Built to train **SciPeerAI-7B** — world's first scientific integrity LLM.

- Live demo: https://scipeerai-ui.vercel.app
- API: https://web-production-f526d.up.railway.app/docs
- GitHub: https://github.com/Abu-Sameer-66/SciPeerAI

## Citation

```bibtex
@dataset{nadeem2026scipeerai_bench,
author = {Nadeem, Sameer},
title = {SciPeerBench: Multi-dimensional Scientific Fraud Detection Benchmark},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Abu-Sameer-66/SciPeerBench}
}
```

## License

CC BY 4.0 — free to use with attribution.