LLM Label Lab

A single-file training simulator for the Datamaker × International Trade Centre-Kenya data annotation programme (Day 7, Generative AI).

Trainees combine answer pieces in the lab on the left and watch a simulated model change on the right. The model is scripted, not a real AI.

Open index.html in a browser. No build step. Works on GitHub Pages.

Class mode (Supabase)

Trainees join with a name, can tap "Ask the trainer", and the trainer sees everyone live, can pause the class and send everyone to a task.

1. Create a Supabase project.
2. Open supabase/setup.sql, replace TRAINER-PIN-HERE with your PIN, and run it in the SQL Editor.
3. Put your Project URL and anon key in the SUPA object in index.html.
4. Deploy (Vercel imports this repo directly). The trainer presses Class, then "Trainer? Enter your PIN".

Without Supabase the page still works on its own; the Class button just stays hidden.

Trainee accounts (Trainee-001 to Trainee-040)

Run supabase/setup2.sql after setup.sql. It creates 40 accounts, each with a personal code. The trainer opens Class, enters the PIN, and presses "Login codes" to download or copy the list. Trainees sign in with their account and code; progress is saved per account and survives a refresh or a different device.

Practice hub, points and challenges

Run supabase/setup3.sql after the other two scripts. Trainees get a "Check my answer" button and hints in the lab, a Practice hub (bot growth stages, learning path, badges, streaks) and a leaderboard with teams of ten. The trainer can hide the leaderboard and start a timed challenge (3, 5 or 10 minutes) from the Class panel.

Project Lab (projects.html)

Eight simulation projects, one page: bounding box (safety equipment, grazing, crop pests), polygon (coffee leaf disease), polyline (lane lines), semantic segmentation (skin lesions), keypoints (exercise coaching) and text sentiment/churn (coffee customers). Trainees label generated scenes, a simulated model learns only from those labels, and a system report shows the real-world consequences on new scenes. Progress is kept in the browser only.

GIGO scenes (gigo/)

The GIGO annotation and simulation app (road, warehouse PPE, drone census, maize scout, LLM answers) lives in gigo/ and runs from the same deployment: open gigo/projects.html. Details, file mapping and tests are in MIGRATION.md. Run its tests with node gigo/tests/run.js and node tests/migration.js. Project Lab source is in src/project-lab/ (python3 build.py rebuilds projects.html).