Voices
On-Device Fine-Tuning Goes Practical: Gemma 270M From 46% to 90% Task Accuracy in 21 Minutes on a Phone
A Google engineer's walkthrough claims you can fine-tune the tiny Gemma 270M model from 46% to 90% task-specific accuracy in 21 minutes directly on a phone, using synthetic task data generation. If it holds up, it's a concrete argument that narrow, on-device specialist models can beat generic large models for bounded tasks at effectively zero inference cost. The claim is single-source social content and unbenchmarked independently, so treat the specific numbers as directional rather than verified.
Source
↳ Follow the thread