This dataset holds recordings from the eye cameras of a Steam Frame headset. The eye images and any health answers are encrypted: only Frametop's maintainer can read them. The rest (where dots were shown, what the eye trackers found, headset motion, fit readings and checklist answers) is readable and licensed under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0). Contributors also grant Frametop's maintainer a separate license to their recordings. By getting access you agree: not to try to decrypt the encrypted files; not to use anything here to identify, recognise, track or profile anyone, or for biometric identification; not to try to find out who contributed; and to delete recordings that a later version of the dataset withdraws. Contributors must be 18 or older, must not live in Illinois, Texas or Washington (USA), and may only record their own eyes. Read the full card before you agree.

Frametop Eyes

Recordings from the infrared eye cameras of the Valve Steam Frame headset, made with Frametop's guided eye recorder. The dataset exists to train an eye-tracking model that runs on the headset itself, for Frametop's gaze pointer, and keeps working where today's trackers struggle: far to the side, through blinks and squints, with glasses or makeup, and when the headset moves on the face.

The eye images are encrypted. Iris images can identify a person, so each contributor's headset encrypts every eye image, and any health answers, to the maintainer's keys before upload. Hugging Face, other contributors and people who download the dataset get only the encrypted files. What's trained from them can be public; the images aren't.

Every contribution comes from a person who recorded their own eyes, reviewed every take, deleted what they didn't want to share, and uploaded it themselves. The maintainer checks each one before it's merged.

Frametop is an independent open-source project. It isn't affiliated with or endorsed by Valve or Hugging Face. Steam and Steam Frame are trademarks of Valve Corporation. The terms here were written in good faith without a lawyer; if something seems wrong, please open a discussion.

Datasheet

Motivation

Why it exists. The headset's own eye tracking and Frametop's classic pupil tracker both lose accuracy at the edges of the view, around blinks, and after the headset shifts on the face. A model trained end to end on the headset's own eye images, from many people and many fits, should close that gap.
Who made it. Frametop's maintainer, DeeJanuz, with recordings from Frametop users who chose to take part.
Funding. None.

Composition

Instances. A round is one recording by one contributor for one bounty (below). It opens with nine dots, then the bounty's steps, grouped into takes.
Cameras. The headset's 2 eye cameras, infrared, 512 x 400, 8-bit grey, about 90 frames a second each: cam0 is the right eye, cam1 the left. They show the eye, lids, lashes and the skin around them, and glasses or contact lenses when worn.
Per take.
- cam0/, cam1/: the frames, one zstd chunk per second, each chunk encrypted (*.zst.age).
- index.tsv: every kept frame: camera, chunk, place in the chunk, and time.
- targets.jsonl: where the dot was (yaw and pitch in degrees from straight ahead, distance), the background's brightness, and the dot's state (look, capture, done).
- prompts.jsonl: what the contributor was asked to do, and when.
- own.jsonl: what Frametop's classic tracker found in each frame (pupil ellipse, glints, gaze).
- steamvr.jsonl: SteamVR's own gaze, as the headset reported it.
- poses.jsonl: the headset's position and rotation.
- ipd.jsonl: the headset's IPD setting.
- take.json: the take's status and the ranges the contributor deleted.
Per round.
- manifest.json: the bounty, the contributor's checklist answers (glasses, contacts or inserts, eye makeup, face pad and strap, a measured IPD if given), software versions, the recipients' public keys, and a summary of the takes.
- fit.json: the fit coach's readings of where each eye sits.
- health.json.age: optional health answers (eye surgery, conditions), encrypted. Only present when the contributor ticked their own consent box.
- SHA256SUMS: over every file, the encrypted ones as they are.
Labels. Where the dot was, at known times, while the contributor looked at it. Training uses the frames where the eyes had settled on a dot. Blink and squint cues are in the prompts.
Size. About 1.2 GB per recorded minute; a round is about 3-5 GB.
Identifiers. Each contributor is a random id (a UUID made on their headset). No names, emails, account names, headset serials or sound are recorded. Round times carry the contributor's time zone. The Hugging Face account that opened a pull request is visible on that pull request, next to the contributor id, but not in the data.

File layout

contributions/<contributor id>/<round YYYYMMDD-HHMMSS-bounty>/
  manifest.json
  fit.json
  health.json.age          (optional)
  SHA256SUMS
  takes/<NN>-<group>/
    cam0/NNNNNN.zst.age    right eye
    cam1/NNNNNN.zst.age    left eye
    index.tsv
    targets.jsonl  prompts.jsonl  own.jsonl  steamvr.jsonl  poses.jsonl  ipd.jsonl
    take.json

The exact formats are in Frametop's gaze/rec/DESIGN.md ("Files"), and the recorder's validate.py checks them.

Collection process

The Frametop eye recorder runs on the headset. A panel inside the headset guides the contributor, who steps through it with the headset's button or the keyboard. Each round:

Put the headset on, comfortably, with a sharp picture. The panel shows the IPD setting.
Nine dots: a 3 x 3 grid to 20 degrees, the round's reference.
The bounty's steps, a few minutes long. The board in bounties.json lists what's needed most: the edges of the view, blinks and lids, headset slips, glasses, jumps and pursuit, depth, light changes, face moves, head turning and reading.

The recorder never uploads anything itself. The contributor reviews every take, can delete any stretch, take or round, exports what's left (the export encrypts the images), and uploads it as a pull request with their own Hugging Face account.

Consent and privacy

Contributors read and accept the consent text in the recorder before their first round, and confirm they're 18 or older and don't live in Illinois, Texas or Washington (USA), whose biometric privacy laws need more than this project can provide yet. The text is in Frametop's repository: gaze/rec/CONSENT.md.
Health answers have their own, separate consent, and are encrypted like the images.
Only the maintainer holds the decryption keys. Decrypted images stay on the maintainer's own computers, for training and checking models, and aren't shared or uploaded anywhere.
Before a pull request is merged, the maintainer checks that every file is complete and that every eye image is encrypted, then decrypts it and looks through it. A pull request that fails is closed and its uploaded files are deleted.
The recordings are used to train and test eye-tracking models. They aren't used to identify people, and the access terms forbid anyone from trying.
Contributors may have further rights over their data depending on where they live (for example under the GDPR); requests go to the maintainer (see Maintenance).

License and contributor grant

The readable files: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Credit contributors by their contributor id, and cite the dataset (below).
The encrypted files are licensed to the maintainer only. Nothing here lets anyone else decrypt or use them.
The contributor grant: each contributor gives Frametop's maintainer (today DeeJanuz, and whoever takes over maintaining Frametop) a non-exclusive, worldwide, royalty-free license to use their contribution for any purpose, commercial included: copying, changing, and training, publishing and selling models made from it. It doesn't include publishing their eye images or health answers.
Privacy isn't licensed. CC BY-NC 4.0 covers copyright and database rights only; the access terms above apply on top of it.
Models trained on this data aren't covered by this card; Frametop's models come with their own license.
See LICENSE.

How to contribute

See CONTRIBUTING.md. In short: install Frametop on a Steam Frame, open the Eye Recorder, record, review, export, then log in and upload on its Upload page. The upload is a pull request from your Hugging Face account.

Withdrawal

A contributor can withdraw at any time, without giving a reason. Before a pull request is merged, close it. After, open a discussion on this dataset, or an issue in Frametop's repository, with your contributor id (shown in the recorder) and which rounds, or "all". Post from the Hugging Face account that uploaded them if you can, or say how the maintainer can check it's you. Don't post anything else that identifies you.

Then, usually within 30 days, the maintainer:

deletes those recordings from the dataset in one commit;
squashes the repository's history, so no older commit holds them, and permanently deletes their stored files, including those in pull-request branches;
deletes their own decrypted copies;
leaves them out of anything trained after that.

This rewrites the repository's history: commit ids change, and anyone who pinned a revision or cloned the repository needs to fetch it again. If you use this dataset, please update your copy when a new version removes recordings, and delete the withdrawn files you hold.

Maintenance

Maintainer: DeeJanuz, through this dataset's discussions or Frametop's issues.
Updates: contributions are merged as they're reviewed. Withdrawals are handled as they arrive.
Errors: report a problem with a recording on the discussions page, quoting the contributor id, round and take.

Known limitations

Few people. Early versions hold recordings from a handful of contributors, all Frametop users. They aren't a balanced sample of eye shapes, ages, eyewear or makeup.
One headset model. Only the Steam Frame's eye cameras.
Labels are dots. Contributors were asked to look at each dot; they don't always do exactly that.
Encrypted. Only the maintainer can train on the images. The readable files alone (dot positions, tracker outputs) are useful for comparing trackers, not for training on images.
Withdrawals. Recordings can disappear in later versions. Pin a revision for reproducibility.

Citation

@misc{frametop_eyes,
  title        = {Frametop Eyes: infrared eye-camera recordings from the Steam Frame},
  author       = {DeeJanuz and the Frametop Eyes contributors},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/DeeJanuz/frametop-eyes}},
  note         = {readable files CC BY-NC 4.0; eye images encrypted}
}

Downloads last month: 27

Total file size:

21.3 kB

Models trained or fine-tuned on DeeJanuz/frametop-eyes

DeeJanuz/frametop-eyes-models

Updated 5 days ago