2796gauravc/functiongemma-pocketassist-v5

FunctionGemma PocketAssist v5

Fine-tuned google/functiongemma-270m-it for a personal notification assistant app. Trained on 300 hand-crafted examples (no synthetic data). Overall accuracy: 100.0% on held-out evaluation set.

Functions

Function	Examples	Accuracy
get_digest	35	100%
get_notes	30	100%
get_password	45	100%
help	30	100%
save_note	40	100%
save_password	40	100%
search	50	100%
status	30	100%

Training Details

• Base model: google/functiongemma-270m-it
• Framework: Unsloth + TRL SFTTrainer
• LoRA: R=16, α=32
• Epochs: 5 | LR: 0.0002 | seq_len: 512
• Key fix: train_on_responses_only — loss only on function call tokens
• packing=False for clean gradients

Android Deployment

See the .task file companion repo for the MediaPipe-ready model. Use with MediaPipe LLM Inference API (minSdk=26).