MicroGPT — In-Browser Lab

A Blazor WebAssembly demo of MicroGPT: train, sample, and anomaly-score a tiny character-level GPT entirely inside the browser. No server, no upload, no API key — after the initial page load the app makes zero network requests (system fonts only, verifiable in the network tab).

What it demonstrates

Tab	Library feature
1 Data	Streaming ingestion with a seeded reservoir sample (--max-docs semantics): memory is bounded by the cap, not the file size, while the character vocabulary and statistics cover the entire file. Plain text or CSV with column selection, length/character charts.
2 Train	Trainer.TrainStep in adaptive chunks on the UI thread with a live loss chart. Editable hyperparameters (embedding dim, heads, layers, steps, lr, seed, char/BPE tokenizer, merges). A held-out slice calibrates loss percentiles (p50/p90/p95/p99) — thresholds are measured, not arbitrary.
3 Generate	Sampler.Sample and prefix-conditioned SampleContinuation, temperature control, download as .txt.
4 Score	Sampler.ComputeLoss per line for anomaly detection: histogram, calibrated percentile threshold slider, out-of-vocab handling, TSV export.
5 Model	Versioned .mgpt binary round trip: download the trained model, upload it later (v1 and v2 files) and skip training entirely.

Project layout

MicroGpt.Web/
  Program.cs, Main.razor          app shell (nameplate, pipeline tabs, footer)
  Models.cs                       records shared by the panels
  Services/
    CsvParser.cs                  RFC4180-ish line parser + delimiter detection
    DatasetService.cs             streaming pass: reservoir sample + full-file stats
    TrainingSession.cs            chunked training, live loss, holdout calibration
    AppState.cs                   session state + file-download helper
  Components/
    Charts/                       dependency-free SVG charts (line, histogram, bars)
    Panels/                       one panel per tab
  wwwroot/                        index.html, css, js interop, bundled sample data

The project references ..\MicroGpt\MicroGpt.csproj and uses only the library's public API — nothing in the library was modified or duplicated.

Add to the solution and run

dotnet sln MicroGpt.sln add MicroGpt.Web
dotnet run --project MicroGpt.Web

Publish

Interpreter mode (no extra workload needed):

dotnet publish MicroGpt.Web -c Release -o publish

Ahead-of-time compilation — strongly recommended for real training speed in the browser (larger download, needs the wasm-tools workload):

dotnet workload install wasm-tools
dotnet publish MicroGpt.Web -c Release -p:RunAOTCompilation=true -o publish-aot

Serve any static file host from publish/wwwroot:

dotnet tool install --global dotnet-serve
dotnet serve -d publish\wwwroot

Deploy to GitHub Pages

1. In wwwroot/index.html change <base href="/" /> to <base href="/MicroGPT/" /> (or your repository name).
2. Use the included workflow: .github/workflows/deploy-web.yml publishes on every push to main. Enable Pages → Source: GitHub Actions in the repository settings.

Notes and limits

• Blazor WebAssembly is single-threaded: training runs in adaptive chunks that yield to the browser (~80 ms per chunk), so the page stays interactive and a Stop button works. Stopping early leaves the learning-rate schedule incomplete.
• Attention cost is quadratic in row length; the app warns when the longest line exceeds 300 characters.
• The .mgpt file stores weights as doubles: parameter count × 8 bytes.
• Everything lives in tab memory. Closing the tab discards the dataset and model — download the .mgpt file to keep it.