ankit-pn/functiongemma-270m-phone-assistant

FunctionGemma 270M: Phone Assistant

This model, developed by ankit-pn, is a full-parameter fine-tune of google/functiongemma-270m-it, specifically engineered to act as a Siri-style phone assistant. It translates user commands into calls for 20 distinct phone tools, covering functionalities like alarms, timers, Wi-Fi, Bluetooth, volume, brightness, weather, calendar, messages, and navigation. For requests not requiring a tool, it provides a direct reply.

Key Capabilities

• Function Calling: Converts natural language commands into structured calls for 20 phone tools.
• Multilingual Support: Trained on a dataset including both English and Hinglish (15% of the training data).
• High Accuracy: Achieves a 75.4% exact match for all functions and arguments on held-out test data, a substantial improvement from the base model's 22.3%.
• Low Parse Errors: Demonstrates a very low rate of parse errors (0.25%).

Good For

• On-device Assistants: Ideal for integrating intelligent function-calling capabilities into mobile applications.
• Voice Control Interfaces: Enabling users to control phone features through natural language commands.
• Research in Function Calling: Provides a strong baseline and a well-documented training process for further research in tool-use models, particularly for low-resource languages like Hinglish.

Limitations

It's important to note that the dataset was LLM-generated, not from real users, and results are from a single seed. Known weaknesses include occasional over-calling on Hinglish small talk, garbling some names, and time conversion inaccuracies. This model is a research artifact, and calls should be validated before execution on a real device.