Tung111/ATIR

License and Redistribution

ATIR is a mixed-license derived dataset. No single license applies to the entire dataset. The licenses and terms of the upstream datasets continue to apply to the corresponding records and derived content.

ATIR-authored contributions

Unless otherwise stated, the annotations, metadata, dataset organization, and newly generated content contributed by the ATIR authors are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). This license does not replace or override the licenses applicable to upstream content.

Upstream datasets

Component	Source	Applicable terms
Speech data derived from LibriSpeech	LibriSpeech	CC BY 4.0
Data derived from SVQ	Google SVQ	CC BY 4.0
CoQA literature and Wikipedia records	stanfordnlp/coqa	CC BY-SA 4.0
CoQA children’s-story records	MCTest	MSR-LA
CoQA examination records	RACE	Non-commercial research use only, subject to the RACE terms
CoQA news records	DeepMind CNN	Apache License 2.0

The CoQA material used in ATIR was obtained from the official stanfordnlp/coqa release, which is itself marked as license: other because it contains records governed by different upstream terms.

Modification notice

ATIR contains modified and derived versions of upstream data. These records are not the original LibriSpeech, SVQ, CoQA, MCTest, RACE, or DeepMind CNN datasets.

The source data were processed and modified by the ATIR authors as part of the ATIR dataset construction pipeline.

For MCTest-derived records, this notice is provided pursuant to the MSR-LA requirement that modified files identify that the data have been changed and state the date of modification.

Redistribution

Redistribution or mirroring is permitted only when the redistributor:

1. Preserves this complete license and provenance notice.
2. Preserves all applicable upstream licenses, attribution notices, and citations.
3. Applies the relevant terms on a component-by-component basis.
4. Distributes MCTest-derived data and derivative works under the same MSR-LA terms and includes the required modification notice.
5. Limits RACE-derived data to non-commercial research use and complies with the RACE terms.
6. Applies CC BY-SA 4.0 to relevant adaptations of the CoQA literature and Wikipedia material.
7. Does not represent the complete ATIR dataset as being governed by a single permissive license.
8. Identifies the official ATIR repository as the canonical source: https://huggingface.co/datasets/Tung111/ATIR

This release does not grant any additional rights to third-party content beyond those provided by the respective upstream licenses.

Citation

Users of ATIR should cite the ATIR paper and the applicable upstream datasets.

@inproceedings{zhao-etal-2026-atir,
    title = "{ATIR}: Towards Audio-Text Interleaved Contextual Retrieval",
    author = "Zhao, Tong  and
      Zhang, Chenghao  and
      Zhu, Yutao  and
      Dou, Zhicheng",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1006/",
    doi = "10.18653/v1/2026.acl-long.1006",
    pages = "22032--22046",
    ISBN = "979-8-89176-390-6",
    abstract = "Audio carries richer information than text, including emotion, speaker traits, and environmental context, while also enabling lower-latency processing compared to speech-to-text pipelines. However, recent multimodal information retrieval research has predominantly focused on images, largely overlooking audio, especially in the setting of interleaved audio-text contextual retrieval. In this work, we introduce the Audio-Text Interleaved contextual Retrieval (ATIR) task, where queries can alternate between audio and text modalities. We construct an ATIR benchmark by integrating several Automatic Speech Recognition (ASR), QA, and retrieval datasets, ultimately unifying four types of contextual retrieval tasks. This benchmark substantially addresses the limitations of existing audio retrieval datasets in semantic retrieval. To study this task, we evaluate several off-the-shelf retrievers and train our ATIR model based on a Multimodal Large Language Model (MLLM). We further propose a novel token compression mechanism, which is orthogonal to existing compression methods, to mitigate the challenge of excessive audio tokens in MLLM-based ATIR models. Experimental results show that our ATIR model achieves significant improvements over strong baselines."
}