# Etho

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Aniket Agrawal](https://cerebralvalley.ai/u/Tekina)
- **GitHub:** https://github.com/Ani0202/etho
- **Demo video:** https://youtu.be/6yoZPROj7PA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2734

Labs that study animal behavior often still score it by hand: someone watches every second of video and logs what the animals do. For social behavior that takes human annotators three to four times the video's length (Segalin et al. 2021, eLife). Behavior scoring underpins neuroscience research and drug-safety studies (ICH S7A; OECD TG 424).

I built Etho, training-free behavior scoring. A researcher uploads lab video, writes behavior definitions in plain English, and gets a scored ethogram, metrics and an auditable methods report. No labeled examples are given to the models and no model is trained.

How it works: Muse Spark picks a phrase for SAM 3.1 to track, and SAM 3.1 follows each animal with a persistent identity. Muse Spark turns each definition into a short Python rule that runs in an isolated sandbox. Muse Spark Open then checks borderline moments from strips of six frames. Every label traces back to the rule and model call that produced it. The researcher can correct a few bouts and have the rule rewritten, so they stay in control.

Results on the public CalMS21 mouse benchmark, all 19 official test videos, scored once: 0.58 macro F1 (averaged over attack, investigation and mount: 0.449, 0.635 and 0.656). Supervised methods trained on labeled data and given pose keypoints reach 0.79-0.86; Etho works from raw video. It scores a video unattended in about its own length (0.88 times real time on test; worst video 1.54 times), at $11.25 per hour of video in model API cost for the fixed benchmark pipeline (excluding the planner and one-time rule writing).

Disclosure: pipeline settings were tuned on 13 labeled development videos. Test-label summaries (behavior and bout counts) informed dev-set selection, the 10 fps sampling rate and the targets; no test predictions were used to choose settings. A second species (fighting fruit flies) runs on the same pipeline but is shown only as a qualitative preview, because it did not clear the confidence bar I set in advance.

Try it without any keys: `make demo` in the repo runs the full pipeline on a 60-second CalMS21 sample clip in about 20 seconds, replaying recorded Meta model responses (clearly labelled as a replay; it never calls a model or the network, and it stops with an error on anything it has no recording for). With your own Meta Model API key, the same pipeline runs live on your video.

Code, results and every disclosure are in the public repo.

Live app, no login: https://etho-vm.tail9158ab.ts.net. Upload a lab video of 2 minutes or less (or the sample clip linked on the page) and get a scored ethogram and report; 10 runs per visitor per day, within a public budget, with precomputed showcase runs always open.

## More from Meta Global AI Developer Hackathon

- [David Koch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2731)
- [wetoyo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2732)
- [Apex Business OS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2733)
- [Engrammo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2735)
- [Mirak](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2736)
- [Verified Asset](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2737)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2734. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt