AirQA: A Comprehensive QA Dataset for AI Research with Instance-Level Evaluation

This repository contains the test set, the metadata, processed_data and papers for the AirQA dataset introduced in our paper AirQA: A Comprehensive QA Dataset for AI Research with Instance-Level Evaluation accepted to ICLR 2026. Detailed instructions for using the dataset will soon be publicly available in our official repository.

AirQA is a human-annotated multi-modal multitask Artificial Intelligence Research Question Answering dataset, which encompasses 1,246 examples and 13,956 papers, aiming at evaluating an agent’s research capabilities in realistic scenarios. It is the first dataset that encompasses multiple question types, also the first to bring function-based evaluation into QA domain, enabling convenient and systematic assessment of research capabilities.

🔍 Quick Start

Load the AirQA dataset in one line using Hugging Face datasets:

from datasets import load_dataset

dataset = load_dataset("OpenDFM/AirQA")

However, we recommend referring to our official repository for complete usage instructions, including the data format and evaluation scripts.

📂 Folder Structure

AirQA
|── data/
|   |── test.parquet # test set (simple, for minimal usage)
|   |── test_data.jsonl # test set (complete, including function-based evaluation)
|   └── uuid2title.json # mapping from paper UUID to title
|── metadata/
|   |── 000ab6db-4b65-5dc0-8393-fbc2c05843c8.json
|   └── ... # more metadata dicts
|── papers/
|   |── acl2016/
|   |   └── 16c3a7ad-d638-5ebf-a72a-bd58f06c16d7.pdf
|   |── acl2019/
|   |   └── c7563d97-695f-5c77-8021-334bf2ff9ddb.pdf
|   |── acl2023/
|   |   |── 001ab93b-7665-5d56-a28e-eac95d2a9d7e.pdf
|   |   └── ... # more .pdf published in ACL 2023
|   └── ... # other sub-folders of paper collections
|── processed_data/
|   |── 000ab6db-4b65-5dc0-8393-fbc2c05843c8.json # cached data for PDF parsing
|   └── ... # more cached data for PDFs
└── README.md

Due to Hugging Face's limit on the number of files in a single folder, we packaged metadata and processed_data into archives.

📊 Dataset Statistics

Our dataset encompasses papers from 34 volumes, spanning 7 conferences over 16 years. The detailed distribution is summarized below.

👇🏻 Click to view the paper distribution of dataset

Folder	Conference	Year	Collected
iclr2024	ICLR	2024	3301
iclr2023	ICLR	2023	31
iclr2020	ICLR	2020	1
neurips2024	NeurIPS	2024	6857
neurips2023	NeurIPS	2023	73
nips2006	NeurIPS	2006	1
acl2024	ACL	2024	161
acl2023	ACL	2023	3083
acl2019	ACL	2019	1
acl2019	ACL	2016	1
emnlp2024	EMNLP	2024	55
emnlp2023	EMNLP	2023	52
emnlp2021	EMNLP	2021	2
emnlp2013	EMNLP	2013	1
icassp2024	ICASSP	2024	18
icassp2023	ICASSP	2023	12
eacl2024	EACL	2024	1
ijcnlp2023	IJCNLP	2023	1
arxiv2025	arXiv	2025	12
arxiv2024	arXiv	2024	53
arxiv2023	arXiv	2023	61
arxiv2022	arXiv	2022	61
arxiv2021	arXiv	2021	43
arxiv2020	arXiv	2020	25
arxiv2019	arXiv	2019	20
arxiv2018	arXiv	2018	11
arxiv2017	arXiv	2017	6
arxiv2016	arXiv	2016	4
arxiv2015	arXiv	2015	1
arxiv2014	arXiv	2014	1
arxiv2013	arXiv	2013	1
arxiv2012	arXiv	2012	1
arxiv2011	arXiv	2011	1
uncategorized	-	-	3
Total	-	-	13956

✍🏻 Citation

If you find this dataset useful, please cite our work:

@misc{huang2025airqacomprehensiveqadataset,
      title={AirQA: A Comprehensive QA Dataset for AI Research with Instance-Level Evaluation}, 
      author={Tiancheng Huang and Ruisheng Cao and Yuxin Zhang and Zhangyi Kang and Zijian Wang and Chenrun Wang and Yijie Luo and Hang Zheng and Lirong Qian and Lu Chen and Kai Yu},
      year={2025},
      eprint={2509.16952},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.16952}, 
}