Dataset Card for TempRAGEval

🤗 Dataset | 📖 arXiv | GitHub

Temporal QA for RAG Evaluation benchmark (TempRAGEval) is a evaluation benchmark of time-sensitive questions. We repurpose two existing datasets:

• TimeQA (Chen et al., 2021)
• SituatedQA (Zhang and Choi, 2021)

We manually augment the test questions with temporal perturbations (e.g., modifying the time period). For example, we change 2019 to 6 May 2021 while preserving the answer Boris Johnson.

In addition, we annotate gold evidence on Wikipedia for more accurate retrieval evaluation. We provide max two gold sentences. If the Wikipedia text chunk contains any one of gold sentences, the text chunk is regarded as gold evidence. It is guaranteed that gold sentences are from the Wikipedia corpus. corpora/wiki/enwiki-dec2021 is the only corpus used in the benchmark, consisting of 33.1M text chunks from ATLAS.

The objective is to evaluate temporal reasoning capabilities and robustness for both retrieval systems and LLMs.

Notes: the samples with empty time_relation are excluded in the evaluation.

You can download this dataset by the following command:

from datasets import load_dataset

dataset = load_dataset("siyue/TempRAGEval")

# print the first example on the test set
print(dataset["test"][0])

Contact

For any issues or questions, kindly email us at: Siyue Zhang (siyue001@e.ntu.edu.sg).

Citation

If you use the dataset in your work, please kindly cite the paper:

@misc{siyue2024mragmodularretrievalframework,
      title={MRAG: A Modular Retrieval Framework for Time-Sensitive Question Answering}, 
      author={Zhang Siyue and Xue Yuxiang and Zhang Yiming and Wu Xiaobao and Luu Anh Tuan and Zhao Chen},
      year={2024},
      eprint={2412.15540},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.15540}, 
}