AuraMind-Maize-Expert-E2B-Finetuned

AuraMind: Fine-Tuned Gemma 3n Maize Expert (E2B)

This is a specialized, fine-tuned version of Google's Gemma 3n model, optimized for diagnosing maize plant health conditions in Nigeria. It was developed as part of the Google - The Gemma 3n Impact Challenge.

Model Derivation

This model was created using an advanced "fine-tune then slice" approach, leveraging the native MatFormer architecture of Gemma 3n:

1. The full unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit model was fine-tuned on a custom dataset of local Nigerian maize diseases.
2. The fine-tuning was performed using LoRA adapters from a hyperparameter sweep, with the champion run (icy-sweep-2) achieving 100% validation accuracy.
3. The trained adapters were merged into the full E4B model.
4. Finally, the E2B sub-model was surgically extracted using the principles from Google's official MatFormer Lab, resulting in this efficient, deployable, and highly accurate expert model.

This process ensures the model has both the high performance of our fine-tuning and the clean, convertible architecture of the official E2B release, making it ideal for on-device deployment with tools like Google AI Edge and MediaPipe.

Project Link: https://github.com/surfiniaburger/AuraMind