FunctionGemma-270M-it-Mobile-Actions-Combined-litert-lm

Function Gemma

Function Gemma is a Google trained light-wight model with 270M parameters to enables users to make connection between their prompt and executive actions. You can find another variant of the model fine-tuned by Google over (google/mobile-actions)[https://huggingface.co/datasets/google/mobile-actions] here: litert-community/FunctionGemma_270M_Mobile_Actions. You can also use this FunctionGemma-270m-it-Extended-Mobile-Actions model variation fine-tuned over extended Mobile Actions tools set dataset in AliRGHZ/Mobile-Actions.

This is the LiteRT version of the fune-tuned model for on-device deployment. For deploying the model into your application, kindly follow the instructions available here. For further fine-tuning you can follow the instructions suggested by Google here.

Notice: Gemma is provided under and subject to the Gemma Terms of Use found at "ai.google.dev/gemma/terms".

Information regarding fine-tuning of the base-model.

Training Configuration:

• Epochs: 2
• Batch size: 4 per device
• Gradient accumulation steps: 8
• Learning rate: 1e-5
• Scheduler: Cosine
• Optimizer: AdamW (fused)
• Precision: bfloat16
• Gradient checkpointing: Enabled
• Completion only loss: True (trains only on model outputs, not prompts)

Training Infrastructure:

Hardware: Google Colab A100 GPU Training time: ~24 minutes for 2 epochs Library versions: transformers==5.2.0 torch==2.10.0 huggingface_hub==1.5.0, trl==0.29.0, accelerate==1.13.0