You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
I have read and agree to the FineVideo Terms of Use reproduced on this dataset card and in THIRD_PARTY_MEDIA.md, including attribution and removal updates, and the LLaVA academic research and education restriction.
Log in or Sign Up to review the conditions and access this dataset content.
ReTurn
When History Helps and Hurts: Selective History Use across Multimodal Turns
Project page · Code and evaluation · Usage guide
ReTurn evaluates selective history use in multimodal conversations: history can supply a relevant question or evidence, while also introducing outdated answers or competing observations.
Current release: Development preview
This repository currently provides 80 Development tasks in 40 matched pairs, an 8-task quick start, and their companion media: 127 video clips and 127 WAV files. This is a preview, not the full benchmark or its Test split. Full dataset: coming soon.
The full benchmark contains 7,000 base tasks: 2,000 Dev tasks (1,500 Dev-Train and 500 Dev-Val) and 5,000 Test tasks. OpenQA and MCQ are two answer views of each task.
Four controlled conditions
| History role | Reference | Conflict |
|---|---|---|
| Task-carrying: earlier question, current media | Reconfirm: answers agree | Reground: the earlier answer is outdated |
| Evidence-carrying: current question, earlier media | Retrieve: weaker current competition | Rebind: competing current evidence |
Matched Single-turn inputs directly supply the target question and media.
Files and use
data/tasks/: task definitions, model inputs, and reference answers.data/preview_manifest.json: the 80-task selection and release version.data/quickstart_manifest.json: the quick-start selection.return-preview-media.zip: companion media, ready for the GitHub installer.data/ATTRIBUTION.json,data/media_manifest.json, anddata/media_checksums.json: provenance, media mapping, and integrity records.
Use the inference adapters and paired evaluation tools in the GitHub repository. This release preserves their native JSONL format; it does not promise automatic loading through datasets.load_dataset.
After accepting the media terms, download return-preview-media.zip. From the ReTurn code repository, run:
python scripts/install_media.py /path/to/return-preview-media.zip --accept-terms
python -m return_eval.check --media
Cached evaluation can run without media or API credentials. New inference requires the media and the appropriate model access. See the linked usage guide for setup and commands.
Access terms — read before requesting access
The following terms are public so that you can read them before accepting access. See also the public media notice.
FineVideo
FineVideo dataset is a collection of over 43.000 YouTube videos. We ask that you read and acknowledge the following points before using the dataset:
- FineVideo is a collection of Creative Commons videos. Any use of all or part of the videos must abide by the terms of the original licenses, including attribution clauses when relevant. We facilitate this by providing provenance information for each data point.
- FineVideo is regularly updated to enact validated data removal requests. By clicking on "Access repository", you agree to update your own version of FineVideo to the most recent usable version specified by the maintainers in the following thread. If you have questions about dataset versions and allowed uses, please also ask them in the dataset's community discussions. We will also notify users via email when the latest usable version changes.
- To host, share, or otherwise provide access to FineVideo, you must include these Terms of Use and require users to agree to it.
LLaVA-Video-178K
The upstream dataset card declares Apache License 2.0 and limits dataset use to academic research and education. We preserve both statements, including the Apache license text and source notices. These statements do not constitute a new license from ReTurn over the original videos. Read the upstream dataset card and Apache license.
By requesting access, you agree to the FineVideo terms above and the stated LLaVA academic-research and education restriction. Hugging Face also asks you to share your username and email with the dataset maintainers as part of gated access.
Sources and licenses
Media are derived from FineVideo and LLaVA-Video-178K. ReTurn task annotations were constructed with large-model assistance and verification as described in the paper; the underlying media are not generated by ReTurn.
Original ReTurn annotations are covered by LICENSE-DATA (CC BY 4.0, limited to rights held by the contributors). Third-party media retain their upstream terms and are not relicensed under that grant. Read THIRD_PARTY_MEDIA.md, including the reproduced FineVideo agreement and LLaVA use restrictions, before downloading. Source attribution and applicable license text accompany the files.
Updates and contact
The current preview is preview-v1. Release changes will be recorded in CHANGELOG.md; task IDs and provenance are retained for traceability. The preview is intended for development and integration, not a substitute for the full Test evaluation.
For technical issues or source-specific removal requests, use GitHub Issues. Include task or media identifiers; do not attach disputed media or private information.
Citation
Paper and citation information will be added when the arXiv record is available.
- Downloads last month
- 28