Score animal behavior from plain-English definitions
Upload lab video, describe each behavior in words, and get a scored ethogram, metrics and an auditable
methods report. Every model call is logged and costed.
Precomputed showcase runs (open instantly, zero spend)
Try it yourself: a 60 s sample clip
Download this clip to see the definitions it is scored with. The whole pipeline takes about 90 s on this clip. The clip is from the dataset's development split.
Clip from CalMS21 (Sun et al. 2021, Caltech), licensed CC BY 4.0
Links: CalMS21 dataset,
CC BY 4.0 licence.
This clip was trimmed to 60 s and re-encoded from a longer CalMS21 recording.
Checking for the sample clip...
Behavior definitions to paste
The dataset authors' definitions (Segalin et al., MARS, eLife 2021). They are already filled in on the upload form;
paste them as they are if you edit the form. Subject: . Partner: .
Unlock live uploads
Scoring your own video makes paid model calls, so it needs the access code from the submission notes.
Browsing the showcase runs above never needs it.
Privacy notice.
Uploads are processed by Meta's Model API in Standard mode. Meta states that it "does not use your Content from
Standard Services to train Meta Models" and keeps API content as described in its Terms and Privacy Policy
(
Meta Model API Data
Commitments). Frame strips from your video are sent to Fireworks (Muse Spark Open) for the tier-1 checks;
Fireworks and Meta keep request logs under their own policies. Within about
24 hours
after upload, the video, its results and every intermediate file made from it are deleted from this server. The
clips uploaded to Meta's Files API for tracking are deleted there through its API right after tracking, again
when the job ends, and again at that deadline if an attempt failed. Kept longer: detector code written from your
behavior definitions (no video or user data; it is reused when someone writes the same definition), the spend
ledger (job id, times, token counts, cost, and the phrase used to find the animals), and server logs that include
job links. Revising a showcase also uses the Standard models. Only upload video you have the right to share.
Detector code runs in a local sandbox. The app estimates the cost before starting and refuses a job that would
pass its spend cap.