Source: Vertex AI — Tune Gemini Models

Google Cloud's overview of supervised, preference and reinforcement tuning for Gemini models.

Contents

Summary

Describes the tuning methods Vertex AI offers for Gemini, which models support each, and guidance on when to tune and how much data to prepare.

Key claims

  • Start with prompting; tune only if quality still falls short, and analyse failures before adding data. Quality beats volume.
  • SFT (input-output pairs; hundreds to thousands) for formats, classification, extraction and style; supported on Gemini 3.5 Flash/Flash-Lite, 3.1 Flash-Lite and 2.5 Pro/Flash/Flash-Lite.
  • Preference tuning (chosen/rejected pairs; hundreds to thousands) on Gemini 2.5 Flash and Flash-Lite.
  • Reinforcement learning fine-tuning (reward functions; dozens to a few hundred prompts) on Gemini 3.5 Flash and 3.1 Flash-Lite, for verifiable multi-step tasks.
  • Checkpoints and continuous tuning (SFT→SFT, SFT→RLFT, RLFT→RLFT) are supported.
  • Parameter-efficient tuning is cheaper and works with less data; full fine-tuning may score higher but costs more to train and serve.

Feeds

Fine-tuning · Preference optimization

Suggest changes →