---
language:
- en
license: cc-by-4.0
size_categories:
- n<1K
task_categories:
- text-classification
dataset_info:
  features:
  - name: title
    dtype: string
  - name: paper_category
    dtype: string
  - name: error_category
    dtype: string
  - name: error_location
    dtype: string
  - name: error_severity
    dtype: string
  - name: error_annotation
    dtype: string
  splits:
  - name: train
    num_bytes: 35801
    num_examples: 91
  download_size: 22781
  dataset_size: 35801
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# SPOT-MetaData

> Metadata & Annotations for **Scientific Paper ErrOr DeTection** (SPOT)   
> *SPOT contains 83 papers and 91 human-validated errors to test academic verification capabilities.*

This dataset is introduced in the paper [When AI Co-Scientists Fail: SPOT-a Benchmark for Automated Verification of Scientific Research](https://huggingface.co/papers/2505.11855).

## 📖 Overview

SPOT-MetaData contains all of the **annotations** for the SPOT benchmark—**no** paper PDFs or parsed content are included here. This lightweight repo is intended for anyone who needs to work with the ground-truth error labels, categories, locations, and severity ratings.

Parse contents are available at: [link](https://huggingface.co/datasets/amphora/SPOT).  
For codes see: [link](https://github.com/guijinSON/SPOT).  

Project page:

> **Benchmark at a glance**  
> - **83** published manuscripts  
> - **91** confirmed errors (errata or retractions)  
> - **10** scientific domains (Math, Physics, Biology, …)  
> - **6** error types (Equation/Proof, Fig-duplication, Data inconsistency, …)  
> - Average paper length: ~12 000 tokens & 18 figures  

## 📜 License
This repository (metadata & annotations) is released under the CC-BY-4.0 license.