functiongemma-mobile-actions-v5-16bit

FunctionGemma Mobile Actions v5

A fine-tuned version of FunctionGemma 270M optimized for mobile device function calling. This model excels at understanding natural language commands and mapping them to structured function calls for common mobile actions.

Model Description

• Base Model: google/gemma-3-1b-it (270M parameters)
• Fine-tuning Method: LoRA (r=128, alpha=128)
• Training Data: google/mobile-actions + synthetic augmentation
• Optimized For: Mobile assistant function calling

Supported Functions

Usage with vLLM

Installation

pip install vllm

Basic Inference

from vllm import LLM, SamplingParams
from datetime import datetime

# Load model
llm = LLM(
    model="essobi/functiongemma-mobile-actions-v5-16bit",
    trust_remote_code=True,
    max_model_len=4096,
)

# Define available tools
tools = [
    {
        "function": {
            "name": "set_alarm",
            "description": "Sets an alarm for a specific time.",
            "parameters": {
                "type": "OBJECT",
                "properties": {
                    "datetime": {"type": "STRING", "description": "The time for the alarm."},
                    "title": {"type": "STRING", "description": "Optional label for the alarm."},
                },
                "required": ["datetime"]
            }
        }
    },
    {
        "function": {
            "name": "create_reminder",
            "description": "Creates a reminder with text and optional time.",
            "parameters": {
                "type": "OBJECT",
                "properties": {
                    "body": {"type": "STRING", "description": "The reminder text."},
                    "datetime": {"type": "STRING", "description": "When to remind."},
                },
                "required": ["body"]
            }
        }
    },
    {
        "function": {
            "name": "send_message",
            "description": "Sends a text message to a contact.",
            "parameters": {
                "type": "OBJECT",
                "properties": {
                    "to": {"type": "STRING", "description": "Contact name or phone number."},
                    "body": {"type": "STRING", "description": "Message content."},
                },
                "required": ["to", "body"]
            }
        }
    },
    # Add more tools as needed...
]

# Build prompt using the training format
def build_prompt(user_input: str, tools: list) -> str:
    now = datetime.now()
    dt_str = now.strftime("%Y-%m-%dT%H:%M:%S")
    day = now.strftime("%A")
    
    # Build function declarations
    func_decls = ""
    for tool in tools:
        func = tool["function"]
        props = func["parameters"].get("properties", {})
        required = func["parameters"].get("required", [])
        
        props_str = ""
        for pname, pinfo in props.items():
            desc = pinfo.get("description", "")
            ptype = pinfo.get("type", "STRING")
            props_str += f"{pname}:{{description:<escape>{desc}<escape>,type:<escape>{ptype}<escape>}},"
        props_str = props_str.rstrip(",")
        
        req_str = ",".join([f"<escape>{r}<escape>" for r in required])
        
        func_decls += f"<start_function_declaration>declaration:{func['name']}{{description:<escape>{func['description']}<escape>,parameters:{{properties:{{{props_str}}},required:[{req_str}],type:<escape>OBJECT<escape>}}}}<end_function_declaration>"
    
    return f"""<start_of_turn>developer
Current date and time given in YYYY-MM-DDTHH:MM:SS format: {dt_str}
Day of week is {day}
You are a model that can do function calling with the following functions{func_decls}<end_of_turn>
<start_of_turn>user
{user_input}<end_of_turn>
<start_of_turn>model
"""

# Generate
prompt = build_prompt("Set an alarm for 7am tomorrow", tools)
sampling_params = SamplingParams(temperature=0.1, max_tokens=150)
outputs = llm.generate([prompt], sampling_params)

print(outputs[0].outputs[0].text)
# Output: <start_function_call>call:set_alarm{datetime:<escape>7am tomorrow<escape>}<end_function_call>

vLLM OpenAI-Compatible Server

# Start the server
python -m vllm.entrypoints.openai.api_server \
    --model essobi/functiongemma-mobile-actions-v5-16bit \
    --port 8000

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

response = client.chat.completions.create(
    model="essobi/functiongemma-mobile-actions-v5-16bit",
    messages=[
        {"role": "user", "content": "Remind me to call the dentist tomorrow"}
    ],
    max_tokens=150,
    temperature=0.1,
)

print(response.choices[0].message.content)

Usage with Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "essobi/functiongemma-mobile-actions-v5-16bit"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

# Use the same prompt building function as above
prompt = build_prompt("What's the weather like?", tools)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.1, do_sample=True)
response = tokenizer.decode(outputs[0], skip_special_tokens=False)

print(response)

Output Format

The model outputs function calls in this format:

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

Parsing Function Calls

import re

def parse_function_call(text: str) -> dict | None:
    """Parse function call from model output."""
    match = re.search(
        r'<start_function_call>call:(\w+)\{([^}]*)\}<end_function_call>', 
        text
    )
    if not match:
        return None
    
    func_name = match.group(1)
    args_str = match.group(2)
    
    # Parse arguments
    args = {}
    for param_match in re.finditer(r'(\w+):<escape>([^<]*)<escape>', args_str):
        args[param_match.group(1)] = param_match.group(2)
    
    return {"name": func_name, "arguments": args}

# Example
output = "<start_function_call>call:set_alarm{datetime:<escape>7am<escape>,title:<escape>Wake up<escape>}<end_function_call>"
parsed = parse_function_call(output)
print(parsed)
# {'name': 'set_alarm', 'arguments': {'datetime': '7am', 'title': 'Wake up'}}

Training Details

• Hardware: 8x Tesla V100-SXM2-32GB
• Training Time: ~48 minutes
• Epochs: 3
• Batch Size: 64 effective (4 per device × 2 grad accum × 8 GPUs)
• Learning Rate: 1e-5 with linear schedule
• Gradient Clipping: max_grad_norm=1.0

Limitations

• Optimized for English only
• Best for single-turn function calling (not multi-turn conversations)
• May struggle with highly ambiguous requests
• Calendar vs Reminder distinction can be tricky for edge cases

License

This model is released under the Gemma License.

Acknowledgments

• Google for the Gemma model family and mobile-actions dataset
• Unsloth for efficient fine-tuning tools