I built a TPU-native medical Q&A fine-tuning pipeline using Gemma 3 with Keras and JAX The project fine-tunes Gemma-3 on medical dialogue da…

I built a TPU-native medical Q&A fine-tuning pipeline using Gemma 3 with Keras and JAX The project fine-tunes Gemma-3 on medical dialogue data from ChatDoctor and evaluates it on MedMCQA, with a larger Gemma 4 model used as a reference baseline. Key results: - Fine-tuned Gemma 3 on TPU - Trained only 2.6M parameters with LoRA - Completed training in 0.7 minutes⚡️ - MedMCQA accuracy stayed flat - Gemma 4 scored better with zero-shot on the same evaluation slice - Fine-tuned outputs became more direct and safety-oriented (win - win) 👉The biggest takeaway: Fine-tuning can improve response behavior without improving benchmark accuracy. Read the complete experiment here: 👇 https://medium.com/google-developer-experts/fine-tuning-gemma-3-on-tpu-for-medical-q-a-with-keras-and-jax-5de4186aba14 #TPUSprint @googledevs @GoogleDevExpert

Source: Francois Chollet (X) | 2026-04-29