independently-platform/functiongemma-tasky-ONNX

FunctionGemma Tasky (ONNX)

This repository contains an ONNX export of functiongemma-tasky, a fine-tuned variant of google/functiongemma-270m-it trained for task/todo function-calling. It targets Transformers.js and includes both full precision and Q4 quantized weights.

Files

• onnx/model.onnx: full-precision weights (fp32)
• onnx/model_q4.onnx: 4-bit quantized weights (q4)

Usage (Transformers.js)

import { pipeline } from '@huggingface/transformers';

// Q4 (smaller, faster)
const pipe = await pipeline('text-generation', 'REPLACE_WITH_HF_REPO', {
  dtype: 'q4',
});

const out = await pipe('Add a task to call Alice tomorrow at 9am', {
  max_new_tokens: 128,
});
console.log(out[0].generated_text);

To load full precision instead:

const pipe = await pipeline('text-generation', 'REPLACE_WITH_HF_REPO', {
  dtype: 'fp32',
});

Transformers.js expects ONNX weights under an onnx/ subfolder, which this repo provides.

Training summary

• Base model: google/functiongemma-270m-it
• Fine-tuning data: synthetic task/todo function-calling prompts, mixed English/Italian, includes user-style typos
• Eval success rate: ~99.5% on a 1500/500 train/eval split

Notes

• Quantized models trade some accuracy for faster inference and smaller size.
• Outputs may not be strict JSON; validate and post-process if needed.