pmarruda/functiongemma-270m-mobile-actions-litertlm-lora

Model Description

This model is a fine-tuned version of functiongemma-270m-it, adapted for the Mobile Actions dataset using LoRA (Low-Rank Adaptation).

The goal of this fine-tuning was to utilize the Mobile Actions feature in Edge Gallery, while keeping training lightweight and practical.

.litertlm file

1. 1.Download the .litertlm file from this repository: https://huggingface.co/pmarruda/functiongemma-270m-mobile-actions-litertlm-lora/blob/main/mobile-actionsq8ekv1024_lora.litertlm
2. 2.Upload it directly to Edge Gallery in the "Mobile Actions" section.
3. 3.Test the different features and enjoy :)

Training Approach

The full notebook, training recipe and raw evaluation outputs are available here: https://github.com/pedroarruda07/functiongemma_fine-tune/tree/main

Instead of following the traditional full fine-tuning recipe, which requires an A100 GPU, this model was fine-tuned using LoRA.

Thanks to LoRA, the entire training process was completed using a single NVIDIA T4 GPU (meaning you can do it for free in Colab 😉), significantly reducing computational cost and energy usage, while achieving very similar performance to the full fine-tuning approach.

This demonstrates that efficient fine-tuning techniques can close most of the performance gap without requiring expensive infrastructure.

Training Results

• —Correct Tool Accuracy: Evaluates if the correct tool was called, even if not with the exact same arguments.
• —Exact Match Rate (Tool + Args): How many examples (eval) got the exact correct function calls. It gives the lower bound of correctness. Some outputs might not match the function arguments exactly but still be acceptable. For example, the show_map function call below:

• —show_map:{'query': 'Maison Marulaz, Besançon, France'}
• —show_map:{'query': 'Maison Marulaz in Besançon, France'}

Edge Deployment

In addition to the standard Hugging Face model files, this repository also provides a `.litertlm` file.

This file is optimized for edge execution and can be run with no additional conversion steps

Follow These Instructions