README.md

---
language:
- en
license: apache-2.0
base_model: google/functiongemma-270m-it
tags:
- function-calling
- mobile-actions
- on-device
- litertlm
- edge-ai
- android
- gemma3
library_name: transformers
pipeline_tag: text-generation
---

# FunctionGemma 270M - Mobile Actions (LiteRT-LM Ready)

A fine-tuned FunctionGemma 270M model optimized for on-device function calling on Android devices using [Google AI Edge Gallery](https://github.com/google-ai-edge/gallery) and LiteRT-LM runtime.

## 🎯 Features

- ✅ **Ready-to-use**: Pre-converted `.litertlm` format for immediate deployment
- ✅ **On-device function calling**: Runs entirely on Android devices without internet
- ✅ **Optimized**: INT8 quantization (~271 MB) for efficient mobile deployment
- ✅ **Mobile Actions**: Supports 6 native Android functions
- ✅ **Low latency**: Optimized with extended KV cache (1024 tokens)

## 📱 Supported Mobile Actions

The model can execute the following Android functions via natural language:

| Function | Example Prompt |
|----------|---------------|
| **Flashlight** | "Turn on the flashlight" |
| **Contacts** | "Create a contact for John Doe with phone 555-1234" |
| **Email** | "Send email to john@example.com" |
| **Maps** | "Show Times Square on the map" |
| **WiFi** | "Turn off WiFi" |
| **Calendar** | "Create a calendar event for Team Meeting tomorrow at 2 PM" |

## 🚀 Quick Start

### Download the Model

```bash
wget https://huggingface.co/Yagna1/functiongemma-270m-mobile-actions/resolve/main/mobile-actions_q8_ekv1024.litertlm
```

Or use Python:

```python
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="Yagna1/functiongemma-270m-mobile-actions",
    filename="mobile-actions_q8_ekv1024.litertlm"
)
print(f"Downloaded to: {model_path}")
```

### Use in Google AI Edge Gallery App

1. **Install** the [Google AI Edge Gallery](https://github.com/google-ai-edge/gallery) Android app
2. **Import** the `mobile-actions_q8_ekv1024.litertlm` file into the app
3. **Navigate** to "Mobile Actions" feature
4. **Test** with natural language prompts like:
   - "Turn on flashlight"
   - "Create contact John Smith"
   - "Show Central Park on map"

## 🏗️ Model Architecture

- **Base Model**: [google/functiongemma-270m-it](https://huggingface.co/google/functiongemma-270m-it)
- **Architecture**: Gemma 3 (270M parameters)
- **Quantization**: INT8 (Dynamic)
- **KV Cache**: Extended to 1024 tokens for longer conversations
- **Runtime**: LiteRT-LM (Google's on-device inference engine)

## 📊 Model Details

| Property | Value |
|----------|-------|
| Parameters | 270M |
| Quantization | INT8 |
| Model Size | 271 MB |
| Format | `.litertlm` (LiteRT-LM) |
| Context Length | 1024 tokens |
| Target Device | Android (ARM) |

## 🔧 Function Calling Format

The model uses LiteRT-LM's native function calling format:

```
<start_function_call>call:function_name{param1:value1,param2:value2}<end_function_call>
```

Example outputs:

**User**: "Turn on the flashlight"  
**Model**: `<start_function_call>call:enableFlashlight{}<end_function_call>`

**User**: "Create contact John Doe with phone 555-1234"  
**Model**: `<start_function_call>call:createContact{contactName:John Doe,phoneNumber:555-1234}<end_function_call>`

## 📚 Training Details

This model was fine-tuned on synthetic Mobile Actions data designed to match LiteRT-LM's expected function calling format. The training focused on:

- Natural language → function call mapping
- Parameter extraction from user queries
- Handling edge cases and variations
- Multi-turn conversation support

## ⚠️ Limitations

- Limited to 6 pre-defined Android functions
- English language only
- Requires Android device with ARMv8-A or newer
- May not handle complex multi-step actions
- Function parameters must match expected schema

## 🤝 Credits

**Original Model**: This is a mirror/re-upload of [JackJ1/functiongemma-270m-it-mobile-actions-litertlm](https://huggingface.co/JackJ1/functiongemma-270m-it-mobile-actions-litertlm)

**Thanks to**:
- **JackJ1** for the original fine-tuning work
- **Google** for FunctionGemma base model and LiteRT-LM runtime
- **Google AI Edge Team** for the Gallery app and tools

## 📄 License

Apache 2.0 (same as base FunctionGemma model)

## 🔗 Resources

- [Google AI Edge Gallery GitHub](https://github.com/google-ai-edge/gallery)
- [FunctionGemma Documentation](https://ai.google.dev/gemma/docs/function_calling)
- [LiteRT-LM Runtime](https://github.com/google-ai-edge/LiteRT-LM)
- [Original Base Model](https://huggingface.co/google/functiongemma-270m-it)

## 📞 Contact

For issues or questions about this model mirror, please open an issue on the [repository](https://huggingface.co/Yagna1/functiongemma-270m-mobile-actions/discussions).

---

**Note**: This model is specifically formatted for the Google AI Edge Gallery app and requires the LiteRT-LM runtime. For general-purpose inference, use the base model or convert to standard formats.