functiongemma-270m-it-mobile-actions

FunctionGemma 270M - Fine-tuned for Mobile Actions

This is a fine-tuned version of google/functiongemma-270m-it specialized for mobile device function calling tasks. The model has been trained on the Google Mobile Actions dataset to reliably execute phone commands through function calling.

Model Details

• Base Model: google/functiongemma-270m-it (270M parameters)
• Fine-tuning Method: Supervised Fine-Tuning (SFT) with TRL
• Task: Function calling for mobile device actions
• Language: English
• License: Apache 2.0

Supported Functions

The model can call the following 7 mobile device functions:

1. turn_on_flashlight - Turns the flashlight on
2. turn_off_flashlight - Turns the flashlight off
3. open_wifi_settings - Opens the Wi-Fi settings
4. send_email - Sends an email (parameters: to, subject, body)
5. create_calendar_event - Creates a new calendar event (parameters: title, datetime)
6. show_map - Shows a location on the map (parameters: query)
7. create_contact - Creates a contact in the phone's contact list (parameters: first_name, last_name, email, phone_number)

Training

Dataset

• Name: Google Mobile Actions Dataset
• Size: ~9,654 examples
• Split: ~8,693 training samples, ~961 evaluation samples

Training Hardware

• GPU: NVIDIA with ~7.65 GB VRAM
• Precision: BF16 mixed precision training

Usage

Installation

pip install transformers torch

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "abubakar-siddik/functiongemma-270m-it-mobile-actions"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

# Define available tools
tools = [
    {
        "function": {
            "name": "create_calendar_event",
            "description": "Creates a new calendar event.",
            "parameters": {
                "type": "OBJECT",
                "properties": {
                    "title": {"type": "STRING", "description": "The title of the event."},
                    "datetime": {"type": "STRING", "description": "The date and time in YYYY-MM-DDTHH:MM:SS format."}
                },
                "required": ["title", "datetime"]
            }
        }
    }
    # Add other tools as needed...
]

# Create messages
messages = [
    {
        "role": "developer",
        "content": "Current date and time: 2024-11-15T05:59:00. You are a model that can do function calling."
    },
    {
        "role": "user",
        "content": "Schedule a team meeting tomorrow at 4pm"
    }
]

# Generate response
prompt = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(response)

Example Output

Input: "Schedule a team meeting tomorrow at 4pm"

Output (function call):

<start_function_call>call:create_calendar_event{title:<escape>team meeting<escape>,datetime:<escape>2024-11-16T16:00:00<escape>}<end_function_call>

Model Performance

The fine-tuned model shows significant improvement over the base FunctionGemma 270M model for mobile action tasks:

• Base Model: Struggles with correct function calling on mobile-specific prompts
• Fine-tuned Model: Reliable function calling with proper parameter extraction

Deployment

On-Device Deployment

This model can be converted to .litertlm format for deployment on Android devices using Google AI Edge tools.

Limitations

• Model is fine-tuned specifically for the 7 mobile actions listed above
• Performance may vary on unseen function combinations
• Temporal reasoning may be limited (e.g., "tomorrow at 4pm" requires date context)
• Model is small (270M parameters) and may have limitations on complex queries

Ethical Considerations

This model should be used responsibly:

• Only execute function calls after user confirmation
• Implement proper authentication and authorization checks
• Be transparent about AI-generated actions
• Respect user privacy for contact and calendar operations

Acknowledgments

• Base model by Google
• Training dataset: Google Mobile Actions
• Training framework: Hugging Face TRL

---

Model Card Version: 1.0.0 Last Updated: 2025-12-29