| --- |
| license: gemma |
| base_model: google/functiongemma-270m-it |
| library_name: transformers.js |
| tags: [function-calling, prism-os, webgpu] |
| --- |
| |
| # Prism OS command center model |
|
|
| FunctionGemma 270M fine-tuned on the commands of |
| [Prism OS](https://github.com/AppleLamps/prism-os), a desktop in the browser whose window |
| manager is a handwired numeric program. It turns a sentence into Prism OS's function |
| calls, or a short reply. It runs in the browser with Transformers.js on WebGPU |
| (`onnx/model_q4f16.onnx`, the graph of |
| [onnx-community/functiongemma-270m-it-ONNX](https://huggingface.co/onnx-community/functiongemma-270m-it-ONNX) with these weights). The |
| PyTorch weights are in `pytorch/`. |
|
|
| ## Results on held-out phrasings |
|
|
| Every eval sentence uses a phrasing template that never appears in training. The |
| checkpoint kept was the best on a separate dev set of eval sentences, which these scores |
| leave out. |
|
|
| | Model | Correct | |
| | --- | --- | |
| | Base FunctionGemma | 108/746 | |
| | Fine-tuned (PyTorch, bf16) | 701/746 | |
| | Fine-tuned, browser build (q4f16), first 100 | 95/100 | |
|
|
| Use is subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). |
|
|
