STOU contains harmful user requests written for AI safety evaluation. By requesting access you agree to use it only for non-commercial research on the safety of AI systems, not to use it to cause harm, and to obtain the seed videos (Ego4D, Aria Everyday Activities, Nymeria+) yourself under their own licenses.

STOU

STOU is a benchmark for evaluating the safety of multimodal assistants on egocentric (first-person) video. Each instance pairs a short egocentric video clip with a harmful user request, given as text and as synthesized speech.

This release contains 1,952 instances built on 713 video clips. The clips come from three seed datasets: Ego4D, Aria Everyday Activities (AEA), and Nymeria+.

Videos are not redistributed. The licenses of the seed datasets do not allow it. Download the seed datasets yourself and run reproduce_clips.py to recreate every clip from the timestamps in clip_timestamps.csv.

STOU/
├── README.md
├── clip_timestamps.csv          # source video + start/end time of each clip (see note below)
├── reproduce_clips.py           # recreates the clips from the seed datasets
└── <HARM_TARGET>/<HARM_TYPE>/
    ├── STOU.json                # instances of this harm target × harm type
    ├── gold_annotations.json    # reference action / harm annotations for evaluation
    └── audio/marin/<instance_id>.wav   # synthesized speech of the text query

There are three harm targets and six harm types, for 18 categories in total.

Harm target	#	Harm type	#
`SELF`	712	`FINANCIAL`	337
`OTHERS`	758	`PHYSICAL`	449
`UNSPECIFIED`	482	`PRIVACY`	198
		`PSYCHOLOGICAL`	345
		`SEXUAL`	282
		`SOCIETAL`	341

Seed dataset	Clips	Instances
Ego4D	485	1,454
Aria Everyday Activities	128	313
Nymeria+	100	185

clip_timestamps.csv lists 872 clips. 713 of them are used by the instances in this release, and the other 159 are not referenced.

Instance format

Each STOU.json maps an instance id to a record with these fields:

Field	Description
`video_name`	Same as `clip_file`.
`source_dataset`	`ego4d`, `aea`, or `nymeria_plus`.
`clip_file`	The clip's file name. It is the key into `clip_timestamps.csv`.
`video_path`	Where `reproduce_clips.py --output-dir clips` writes the clip, relative to the release root.
`video_description`	Model-generated description of the clip that was used during query generation. It is `null` for AEA clips (not released, per the AEA license) and for instances created without a generated description.
`text-query`	The harmful user request (text).
`audio`	Speech version of `text-query` (24 kHz mono WAV). The path is relative to the release root.

gold_annotations.json in the same folder uses the same instance ids. It holds the reference annotations used for evaluation:

Field	Description
`axis1`, `axis2`	Harm target and harm type.
`text_query`	Same as `text-query` above.
`gold_action`	`intended_action`: the action the user is asking for.
`gold_harm`	`main_harm` (`potential_target`, `harm_mechanism`, `consequence`) and `acceptable_alternate_harms`.
`gold_annotation_status`	Annotation status (`ok` for all released instances).
`video_name`, `source_dataset`, `clip_file`, `video_path`	Same as in `STOU.json`.

Reproducing the video clips

1. Get the seed datasets

Accept each dataset's license and download it from the official source.

Dataset	What to download	Expected layout (`--*-dir`)
Ego4D	v2 `full_scale` videos	`<dir>/<video_uid>.mp4`
Aria Everyday Activities	release 1.0.0, `main_vrs`	`<dir>/<sequence>/recording.vrs`
Nymeria+	head recording VRS	`<dir>/<sequence>/recording_head/data/data.vrs`

The aria_dataset_downloader from projectaria_tools creates the AEA layout above.

2. Run the script

pip install projectaria-tools==2.2.0 moviepy==2.2.1 numpy tqdm   # ffmpeg must be on PATH

python reproduce_clips.py \
    --timestamps clip_timestamps.csv \
    --output-dir clips \
    --ego4d-dir   /path/to/ego4d/v2/full_scale \
    --aea-dir     /path/to/aea \
    --nymeria-dir /path/to/nymeria_plus \
    --num-workers 8

The script writes clips to clips/<source_dataset>/<clip_file>, which matches video_path in the JSON files. You can omit a --*-dir flag to skip that dataset.

Timestamp format (`clip_timestamps.csv`)

Column	Description
`source_dataset`, `source_id`	Seed dataset, and its Ego4D `video_uid` or Aria sequence name.
`clip_file`	Clip file name, as referenced by the JSON files.
`start_sec`, `end_sec`, `duration_sec`	Clip window in seconds.
`time_reference`	What `start_sec` is measured from. For Ego4D this is the start of the `full_scale` video. For AEA and Nymeria+ it is the first audio record (stream `231-1`, RECORD_TIME), which is the origin used by `projectaria_tools`' `Vrs2Mp4Converter`.
`time_origin_ns`, `clip_start_device_ns`, `clip_end_device_ns`	AEA / Nymeria+ only: the same window as absolute device timestamps (ns).

Licenses and attribution

The seed videos remain under their original licenses. You must obtain them yourself and comply with each license. Use is limited to non-commercial research.

Ego4D: Ego4D License Agreement. Please cite: "Egocentric Live 4D Perception (Ego4D) Database: A large-scale first-person video database, supporting research in multi-modal machine perception for daily life activity. Ego4D Consortium 2020. https://sites.google.com/view/ego4d/home".
Aria Everyday Activities: AEA Dataset License Agreement (non-commercial research only). Attribution: "Aria Everyday Activities Dataset, Meta Reality Labs-R".
Nymeria+: non-audio data under CC BY-NC 4.0; audio data under the Meta Non-Commercial Research License. Copyright (c) Meta Platforms, Inc. and affiliates.

The STOU text queries, labels, gold annotations, and synthesized audio are released under CC BY-NC 4.0.

Citation

@misc{han2026safetywhombenchmarkingunderstanding,
      title={Safety for Whom? Benchmarking Understanding Beyond Refusal in Egocentric Multimodal Assistants}, 
      author={Yewon Han and Jisu Jeon and Suhyun Choi and Juan Yeo and Minsoo Jo and Gihoon Kim and Hyunggyu Lim and Taesup Kim},
      year={2026},
      eprint={2610.17728},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.17728}, 
}

Downloads last month: 13

Total file size:

551 MB

Paper for Chloeing/STOU

Safety for Whom? Benchmarking Understanding Beyond Refusal in Egocentric Multimodal Assistants

Paper • 2610.17728 • Published 12 days ago

Access STOU