LUMON
Macrodata Refinement · Behavior Labeller

The work is mysterious and important.

refinement in progress
This is a local-server tool, not a hosted page. Unlike the other vibes, the MDR Labeller runs a small Python backend (server.py) that renders behavior clips on demand from lab data on /snlkt. It can't run on vibes.tlab.sh — clone the repo and run it locally (a self-contained demo mode needs no lab data). See the quick start below.

01What it is

A Severance / Lumon "Macrodata Refinement"-styled web front-end for labelling animal social-behavior clips. A dense grid of rounded clip tiles sits where the show's floating numbers do: hover a tile to enlarge it and play a rank-colored skeleton clip, then drag it into one of the numbered bins (the ethogram) to label it — or press the bin's number key while hovering. The bins fill and the completion bar advances as you refine, MDR-style.

It's a drop-in alternative to a lab keyboard-driven labelling GUI: same inputs, same labels.csv schema, so the two share resume state.

02How refinement works

03Quick start

# clone the vibes repo, then:
cd mdr-labeller

# demo mode — synthetic clips, no lab data, needs only numpy + ffmpeg
python3 server.py --source demo

# then open the printed URL (default http://127.0.0.1:8752/)

Real mode reads a pre-rendered behavior-clip corpus and writes labels to a labels.csv; it needs numpy, opencv-python, pandas, and the ffmpeg CLI. A one-shot launcher (./run.sh start|status|stop|restart) runs the server detached and prints the ssh -L … tunnel line for reaching it from your laptop. Full details, config keys, and how to plug in a different clip source are in the README.