do you really need that much data for fine-tuning?
The actual plan was to distill Gemini 2.5 Pro into the best on-device voice dictation model.
Pretty sure it would have worked. Alas.