dousery/functiongemma-mobile-actions

Model Card

This model is fine-tuned version of google/functiongemma-270m-it model for mobile action function calling tasks.

Intended Use

Handles function-calling style mobile actions such as creating calendar events, sending emails, adding contacts, showing maps, managing Wi‑Fi, and toggling the flashlight, based on the google/mobile-actions dataset.

Model Details

• Base Model: google/functiongemma-270m-it
• Fine-tuning: SFT with LoRA
• Dataset: google/mobile-actions
• Params: ~270M (base) + LoRA merged into final weights

Quick Start

pip install torch transformers datasets accelerate huggingface_hub

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "dousery/functiongemma-mobile-actions"  
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    device_map="auto" if device == "cuda" else None,
    trust_remote_code=True,
).eval()

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
if device == "cpu":
    model = model.to(device)

dataset = load_dataset("google/mobile-actions", split="train")
text = tokenizer.apply_chat_template(
    dataset[0]["messages"][:2],
    tools=dataset[0]["tools"],
    tokenize=False,
    add_generation_prompt=True,
).removeprefix("<bos>")

inputs = tokenizer(text, return_tensors="pt").to(device)
with torch.no_grad():
    _ = model.generate(
        **inputs,
        max_new_tokens=256,
        streamer=TextStreamer(tokenizer, skip_prompt=True),
        top_p=0.95,
        top_k=64,
        temperature=1.0,
    )

Training Summary

• Frameworks: Unsloth + TRL, PyTorch 2.9.1, Transformers 4.57.3
• Steps: 100 (SFT with LoRA, then merged)
• Effective Batch Size: 8 (bs=4, grad accum=2)
• LR / Scheduler: 2e-4, linear
• LoRA: r=16, alpha=16, dropout=0, ~3.8M trainable params
• Seq Len: 4096
• Hardware: NVIDIA H100 80GB on Modal
• Final Train Loss: 0.2408 | Eval Loss: ~0.0129

Limitations

• Trained for only 100 steps; niche mobile-action domain.
• Datetime formats can drift slightly.
• Best on GPU for speed; CPU works but slower.

Citation

@misc{functiongemma-mobile-actions,
  title={FunctionGemma Mobile Actions - Merged for Mobile Function Calling},
  author={dousery},
  year={2025},
  howpublished={\url{https://huggingface.co/dousery/functiongemma-mobile-actions}}
}

License

Apache-2.0 (inherits base model license).