# Model playground · gpu-engine

[Skip to content](#docs-content)

# Model playground

Choose an example and edit its `.model.ts`, `.eval.ts` or `decode.ts` file. The preview typechecks your edits against gpu-engine's type declarations and reports any TypeScript errors.

- ## `time-tags`

  Tags every token in a short English time phrase (such as "tomorrow at 5 pm") as `day`, `hour` or `none`.

  **Try:** Remove `'day'` from `head.labels` and check the TypeScript error in `decode.ts`.
- ## `intent`

  Classifies a short request (such as "book a table for two") into one of four intent labels using a `pooled` head.

  **Try:** Add `'refund'` to `head.labels` and see `decode.ts` require a route for it.
- ## `color`

  Predicts three numeric RGB channels from 0 to 1 for a color description (such as "deep sea green") using a `pooled` values head.

  **Try:** Swap `outputs: 3` for `labels: ['red', 'green', 'blue']` and watch `decode.ts` fail typechecking.

This playground typechecks the example files in your browser against gpu-engine's type declarations. It does not run `defineModel` or `plan`, does not train because the examples ship no datasets or generators, and never runs `promote`.

[Check project status](https://gpu-engine.dev/docs/start/status)

## Training on your machine

All six CLI commands (`new`, `loop`, `promote`, `parity`, `doctor` and `inspect`) run locally in Node today. `loop` checks data splits, trains with int6 quantization-aware training, scores on dev and logs each try.

[Read Train and tune](https://gpu-engine.dev/docs/build/train-and-tune#the-dev-tune-loop)

## Sealed test set

The playground never runs `promote`, and none of its examples include a `data/` or `data/test/` directory. Because no test data is ever bundled into the browser, there is nothing to read, count or leak.

Real model packages keep `data/test/` out of the tuning loop through the harness enforcement levels. A browser cannot hold a level 3 secret because anything delivered to a client can be inspected.

[Read about enforcement levels](https://gpu-engine.dev/docs/build/evaluate-and-promote#enforcement-levels)
