BLOG · GOOGLE CLOUD · GEMINI · FINE-TUNING
Gemini fine-tuning for classification: tune it on your team's decisions
How supervised fine-tuning of Gemini works on Google Cloud (Vertex AI, now Agent Platform), what it costs, and how to tune it from a Google Sheet.
Fine-tuning and the StayCharted workflow
Your team has categorized thousands of records: tickets, emails, claims, documents. If your company runs on Google Cloud, you can tune Gemini on those decisions in your own project, so it categorizes the next ones the same way.
A note on names: Gemini is tuned on Gemini Enterprise Agent Platform, which Google called Vertex AI until April 2026. The console now says "Agent Platform"; many people still search for Vertex AI, so this post uses both.
What tuning is
A base Gemini model knows language in general. Supervised fine-tuning trains it further on your examples, each one an input and the answer you want, so it gives your answers without being told every time. For classification, each example is a record and the category your team gave it. Google trains a small adapter on top of Gemini (the "adapter size" setting), and you get a tuned model with its own endpoint in your project.
Prompting or tuning?
| Prompting Gemini | Tuning Gemini | |
|---|---|---|
| What you give it | Instructions and a few examples, on every call | Hundreds or thousands of examples, once |
| Good for | A few categories that are easy to describe | Many categories, fine distinctions, your team's judgment calls |
| Cost per call | Longer prompts, more tokens | Short prompts, since the examples are built in |
| Effort | Write and maintain a prompt | Prepare data, tune, test, keep it current |
How Gemini tuning works
- Prepare a JSONL file: one example per line, a user turn with the record and a model turn with the category, plus an optional system instruction. One malformed line fails validation.
- Upload it to Cloud Storage, with a validation set held back (10 to 20% is common).
- Create a tuning job on a current Gemini model that supports tuning, such as Gemini 3.5 Flash (Google calls the 2.5 models out of date), choosing epochs, adapter size and learning rate multiplier, or keeping the defaults.
- Wait for the job. It produces a tuned model and an endpoint in your project.
- Call the endpoint instead of the base model, with the same instructions you tuned with, and test it on examples it never saw.
What it costs
On your Google Cloud bill:
- Tuning: a one-time charge by tokens processed (examples × tokens × epochs).
- Answers: per token, at the applicable tuned-model inference rate.
Classification records are short, so tuning a few thousand of them is usually a small bill. Check Google's pricing page for your model and region.
The parts that take the time
The tuning job is the easy part. The time goes on everything around it: turning a spreadsheet into valid JSONL, holding back honest test rows, service accounts and bucket permissions, waiting on the job, testing every held-out row, working out accuracy per category, deciding which answers a person should check (Gemini gives an answer, not a confidence), and doing it all again when your categories change.
Tune Gemini in your own project, from a spreadsheet or Google Sheet
StayCharted runs the whole loop in your Google Cloud project, and if your team's history is already in Google Sheets, it starts from the sheet:
- Connect once: a bucket, a service account with Agent Platform User on the project and Storage Object Admin on the bucket, and its key, in Settings → AI platforms → Google Cloud with Also train models here. The guide has every click.
- Send your examples: from Google Sheets with the StayCharted add-on (Extensions → StayCharted), or upload a spreadsheet.
- Choose Dedicated AI Model → Where it trains: Google Cloud and a Gemini base, such as Gemini 3.5 Flash.
- StayCharted does the work: holds back test rows, writes the JSONL with the same instructions it will ask with, puts it in your bucket, starts the tuning job, deletes the file when it ends, and tests the tuned model on the held-out rows.
- Read the report and publish. The tuned Gemini answers from your project, per token, with no hourly hosting fee.
- Use it: fill the empty rows of the same sheet, or use it from the API, Zapier, Power Automate, or Claude and ChatGPT. Unsure answers go to the Review Queue.
Not ready to tune? StayCharted can also use Gemini without tuning it ("Use your own AI, without training"), with your categories, your labeling guide and the most similar past examples in every question. Test both on the same rows and publish the better one.
Provider documentation and pricing
Check the rate for your tuned model
Google Cloud bills training, inference and storage. Tuned-model inference pricing depends on the model: Google lists a 1.5× base inference rate for Gemini 3 and later tuned models, while older Gemini tuned models retain the base prediction rate. Check the linked pricing page before starting. StayCharted-managed cloud training requires Business.
Common questions about Gemini fine-tuning (Vertex AI)
How do I fine-tune Gemini?
Prepare a JSONL file of examples, upload it to Cloud Storage, create a supervised tuning job under Agent Platform → Models → Tuning, then call the tuned model's endpoint and test it on held-out examples.
Is Vertex AI still called Vertex AI?
Google renamed it Gemini Enterprise Agent Platform in April 2026. Tuning carried over.
How much does Gemini fine-tuning cost?
Tuning is charged by tokens processed; a tuned model answers at the applicable tuned-model inference rate, with no hourly hosting fee. Check Google Cloud's pricing page for current rates.
Can StayCharted fine-tune Gemini in my project?
Yes, on Business: Gemini 3.5 Flash, and the 2.5 models in projects Google still lets use them, from a spreadsheet or straight from Google Sheets.
Does my data leave my Google Cloud project?
The training file goes to your bucket and is deleted when tuning ends; the tuned model stays in your project and answers there. StayCharted holds your examples to build the file and test the model, as it does for any model.