metadata
license: cc-by-sa-4.0
language:
- en
tags:
- citation
- scientific paper
- grounding
- source attribution
- paper citations
- citation benchmark
- text benchmark
CiteME is a benchmark designed to test the abilities of language models in finding papers that are cited in scientific texts.#
Dataset Structure
The dataset is provided in CSV format and includes the following columns:
| Column Name | Description |
|---|---|
id |
A unique identifier for each paper, used consistently across all experiments. |
excerpt |
The text excerpt from the source paper that describes the target paper. |
target_paper_title |
The title of the paper that is being cited in the excerpt. |
target_paper_url |
The URL linking to the target paper. |
source_paper_title |
The title of the paper from which the excerpt is taken. |
source_paper_url |
The URL linking to the source paper. |
year |
The publication year of the source paper. |
split |
Indicates the dataset split: train or test. |
Example
| id | excerpt | target_paper_title | target_paper_url | source_paper_title | source_paper_url | year | split |
|---|---|---|---|---|---|---|---|
| 1 | "As demonstrated in [Smith et al., 2020], the proposed method improves accuracy significantly." | "Improving Accuracy in ML Models" | https://example.com/target1 | "Advancements in Machine Learning" | https://example.com/source1 | 2020 | train |
| 2 | "Building upon the framework introduced by [Doe, 2019], we extend the applicability to NLP tasks." | "Framework for NLP Applications" | https://example.com/target2 | "Foundations of NLP" | https://example.com/source2 | 2019 | test |
Load the Dataset
You can load the dataset using popular data processing libraries such as pandas.
import pandas as pd
dataset = pd.read_csv('DATASET.csv')
print(dataset.head())
If you find our work helpful, please use the following citation:
@misc{press2024citeme,
title={CiteME: Can Language Models Accurately Cite Scientific Claims?},
author={Ori Press and Andreas Hochlehnert and Ameya Prabhu and Vishaal Udandarao and Ofir Press and Matthias Bethge},
year={2024},
eprint={2407.12861},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2407.12861}
}