README.md

---
title: FunctionGemma 270M Mobile Actions SFT
emoji: 📱
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
---

# FunctionGemma 270M — Mobile Actions SFT

Full fine-tune of `google/functiongemma-270m-it` for function calling on the
`google/mobile-actions` dataset, using TRL's `SFTTrainer`
(prompt/completion format, `completion_only_loss=True`).

## Setup

1. Accept the licenses for both gated repos:
   - https://huggingface.co/google/functiongemma-270m-it
   - https://huggingface.co/datasets/google/mobile-actions
2. Add your token as a Space secret named `HF_TOKEN` (Settings -> Secrets).
3. Use a **GPU** Space (T4/L4/A10G or better). On CPU, training and scoring are slow —
   shrink the train/eval subsets.

## Why gradio is pinned to 5.x

`gradio` 6.x requires `huggingface_hub >= 1.2.0`, but `transformers==4.57.1`
requires `huggingface_hub < 1.0`. Pinning the SDK to `5.49.1` lets pip resolve a
`huggingface_hub` 0.x that satisfies both.