When History Helps and Hurts: Selective History Use across Multimodal Turns
Paper • 2610.11948 • Published
Datasets:
You need to agree to share your Hugging Face 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 · Paper · Code and evaluation · Usage guide
ReTurn evaluates selective history use: conversational history can supply a relevant question or the evidence needed by a current request, while also introducing outdated answers or competing observations. Each evaluated model generates its own historical replies.
Full release: return-full-v1
| Split | Base tasks | Operation pairs |
|---|---|---|
| Dev-train | 1,500 | 750 |
| Dev-val | 500 | 250 |
| Test | 5,000 | 2,500 |
| Total | 7,000 | 3,500 |
Each task has OpenQA and four-option MCQ views: 14,000 task/interface views, not independent tasks. Short and Long contain two and five media-bearing user turns, including the final turn; they are not universally item-matched. The 80-task Development preview and eight-task quickstart remain available as selections.
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| Task-carrying: earlier question, current media | Reconfirm: answers agree | Reground: the earlier answer is outdated |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 |
Each local pair preserves the target question, media and answer. Matched Single-turn inputs directly supply the target question and media. Visual/audio labels identify necessary evidence; backend input channels follow the evaluation protocol.
Download and use
Use the updated ReTurn code, including its full-release SHA256 reference. After accepting the access terms, download this repository with its folder structure intact. Media ZIPs are independent shards: keep them under media/ rather than extracting them manually.
From the updated code checkout:
python -m return_eval.install_data --dataset /path/to/return-data
python -m return_eval.check --manifest data/test_manifest.json --media
Configure model and judge access as described in the usage guide. Inference defaults to the eight-task quickstart. To evaluate the formal Test split:
python -m return_eval.run --config configs/qwen38_api.json \
--manifest data/test_manifest.json --out runs/qwen38_test
python -m return_eval.score --run runs/qwen38_test --judge
python -m return_eval.report --run runs/qwen38_test
Use --dry-run on the run command to validate inputs without model calls. API inference and judging use your own credentials and may incur charges. Dev-train and Dev-val have separate manifests and are not included in formal Test results. The native JSONL format uses the supplied installer, not an assumed datasets.load_dataset interface.
Files and integrity
- data/tasks/: tasks, frozen model-visible requests and scoring references.
- data/*_manifest.json: split, preview and quickstart selections.
- data/media_manifest.json and data/ATTRIBUTION.json: source and attribution records.
- data/media_files.json: installed paths and sizes.
- data/full_media_checksums.json: SHA256 for installed media and ZIP shards; identical to the code reference.
- media/: 14 independent ZIP shards, about 12.55 GiB total (13.39 GiB installed).
- checksums/SHA256SUMS: downloaded-file hashes, including all shards and published data/documents; paths are relative to the repository root.
On macOS/Linux, verify the download from the dataset root with shasum -a 256 -c checksums/SHA256SUMS. The installer verifies shard and extracted-media hashes. The release contains 1,283 clip/audio pairs represented by 48,686 physical files, including per-task copies and legacy paths; file copies are not unique source samples.
The original return-preview-media.zip remains a separate legacy download for the existing preview workflow. Full installation uses media/ shards and does not require that old archive. Preview results are not a reproduction of the full Test leaderboard. See DATA_FORMAT.md and RELEASE_NOTES.md.
Scoring
The evaluator retains strict and content scores. Rules match frozen gold/aliases and parse MCQ responses; unresolved responses are reviewed by the configured judge. The paper scorer uses gpt-5.6-sol with medium reasoning effort. Equivalent answers receive content credit; incorrect or uncertain answers do not. Unreviewed responses remain pending, and failed requests are excluded explicitly. Full model outputs and judge caches are not included in this data release.
Access terms — read before requesting access
Original task annotations are CC BY 4.0, limited to rights held by ReTurn contributors. Third-party media retain their upstream conditions. Dataset construction uses large-model-assisted annotation and verification; source media are not generated by ReTurn.
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, the license text and source notices. These do not constitute a new license from ReTurn over original videos.
By requesting access, you agree to the FineVideo terms above and the stated LLaVA restriction. See THIRD_PARTY_MEDIA.md and the public media notice. Hugging Face also requests account/contact sharing as part of gated access.
Updates and citation
See CHANGELOG.md for version history. Task IDs and provenance are retained. Technical issues and source-specific removal requests can be submitted through GitHub Issues.
Please cite the ReTurn paper. Use the arXiv citation export for author and version metadata.
- Downloads last month
- 53