Interactive React app for visualizing small random MLPs and estimator behavior.
Live demo: https://aicrowd.github.io/whestbench-explorer/
WhestBench Explorer is the educational and debugging companion to whestbench (the library + CLI). It is not the submission interface — official local scoring lives in whest run --estimator <path> --runner server in the main repo.
All computation runs in your browser (Web Workers + TensorFlow.js); there is no backend.
npm install
npm run devOpen http://localhost:5173.
| Panel | Description |
|---|---|
| Network Graph | Layer/neuron structure with value-oriented coloring |
| Signal Heatmap | Layer-wise activation behavior |
| Estimator Comparison | Per-layer error across available estimators |
| Controls | Width, depth, seed, and budget knobs |
- Reproduce a pattern on small MLPs.
- Inspect where errors spike by depth.
- Convert that intuition into estimator logic.
- Re-test with the official local scorer in whestbench.
| Command | Purpose |
|---|---|
npm run dev |
Start the Vite dev server |
npm run build |
Build a static bundle into dist/ |
npm run preview |
Preview the production build locally |
npm run lint |
Run ESLint |
npm run test |
Run the Vitest suite |
Pushes to main trigger .github/workflows/deploy.yml, which builds with Vite and deploys dist/ to GitHub Pages at https://aicrowd.github.io/whestbench-explorer/.
The Vite base is set to /whestbench-explorer/ so the bundle resolves correctly under the project Pages path.