About supervised fine-tuning for Gemini models

Supervised fine-tuning is a good option when you have a well-defined task with available labeled data. It's particularly effective for domain-specific applications where the language or content significantly differs from the data the large model was originally trained on. You can tune text, image, audio, and document data types.

Supervised fine-tuning adapts model behavior with a labeled dataset. This process adjusts the model's weights to minimize the difference between its predictions and the actual labels. For example, it can improve model performance for the following types of tasks:

  • Classification
  • Summarization
  • Extractive question answering
  • Chat

For a discussion of the top tuning use cases, check out the blog post Hundreds of organizations are fine-tuning Gemini models. Here's their favorite use cases.

To learn more, see When to use supervised fine-tuning for Gemini.

Supported models

The following Gemini models support supervised fine-tuning:

For models that support thinking, we suggest setting the thinking budget to off or its lowest value. This can improve performance and reduce costs for tuned tasks. During supervised fine-tuning, the model learns from the training data and omits the thinking process. Therefore, the resulting tuned model can perform tuned tasks effectively without a thinking budget.

Limitations

Gemini 2.5 Flash
Gemini 2.5 Flash-Lite

Specification Value
Maximum input and output training tokens 131,072
Maximum input and output serving tokens Same as base Gemini model
Maximum validation dataset size 5000 examples
Maximum training dataset file size 1GB for JSONL
Maximum training dataset size 1M text-only examples or 300K multimodal examples
Adapter size Supported values are 1, 2, 4, 8, and 16

Gemini 2.5 Pro

Specification Value
Maximum input and output training tokens 131,072
Maximum input and output serving tokens Same as base Gemini model
Maximum validation dataset size 5000 examples
Maximum training dataset file size 1GB for JSONL
Maximum training dataset size 1M text-only examples or 300K multimodal examples
Adapter size Supported values are 1, 2, 4, and 8

Gemini 2.0 Flash
Gemini 2.0 Flash-Lite

Specification Value
Maximum input and output training tokens 131,072
Maximum input and output serving tokens Same as base Gemini model
Maximum validation dataset size 5000 examples
Maximum training dataset file size 1GB for JSONL
Maximum training dataset size 1M text-only examples or 300K multimodal examples
Adapter size Supported values are 1, 2, 4, and 8

Known issues

  • Applying controlled generation when submitting inference requests to tuned Gemini models can result in decreased model quality due to data misalignment during tuning and inference time. During tuning, controlled generation isn't applied, so the tuned model isn't able to handle controlled generation well at inference time. Supervised fine-tuning effectively customizes the model to generate structured output. Therefore you don't need to apply controlled generation when making inference requests on tuned models.

Use cases for using supervised fine-tuning

Foundation models work well when the expected output or task can be clearly and concisely defined in a prompt and the prompt consistently produces the expected output. If you want a model to learn something niche or specific that deviates from general patterns, then you might want to consider tuning that model. For example, you can use model tuning to teach the model the following:

  • Specific structures or formats for generating output.
  • Specific behaviors such as when to provide a terse or verbose output.
  • Specific customized outputs for specific types of inputs.

The following examples are use cases that are difficult to capture with only prompt instructions:

  • Classification: The expected response is a specific word or phrase.

    Tuning the model can help prevent the model from generating verbose responses.

  • Summarization: The summary follows a specific format. For example, you might need to remove personally identifiable information (PII) in a chat summary.

    This formatting of replacing the names of the speakers with #Person1 and #Person2 is difficult to describe and the foundation model might not naturally produce such a response.

  • Extractive question answering: The question is about a context and the answer is a substring of the context.

    The response "Last Glacial Maximum" is a specific phrase from the context.

  • Chat: You need to customize model response to follow a persona, role, or character.

You can also tune a model in the following situations:

  • Prompts are not producing the expected results consistently enough.
  • The task is too complicated to define in a prompt. For example, you want the model to do behavior cloning for a behavior that's hard to articulate in a prompt.
  • You have complex intuitions about a task that are difficult to formalize in a prompt.
  • You want to reduce the context length by removing the few-shot examples.

Configure a tuning job region

User data, such as the transformed dataset and the tuned model, is stored in the tuning job region. During tuning, computation could be offloaded to other US or EU regions for available accelerators. The offloading is transparent to users.

  • If you use the Vertex AI SDK, you can specify the region at initialization. For example:

    import vertexai
vertexai.init(project='myproject', location='us-central1')

  • If you create a supervised fine-tuning job by sending a POST request using the tuningJobs.create method, then you use the URL to specify the region where the tuning job runs. For example, in the following URL, you specify a region by replacing both instances of TUNING_JOB_REGION with the region where the job runs.

     https://TUNING_JOB_REGION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/TUNING_JOB_REGION/tuningJobs

  • If you use the Google Cloud console, you can select the region name in the Region drop-down field on the Model details page. This is the same page where you select the base model and a tuned model name.

Quota

Quota is enforced on the number of concurrent tuning jobs. Every project comes with a default quota to run at least one tuning job. This is a global quota, shared across all available regions and supported models. If you want to run more jobs concurrently, you need to request additional quota for Global concurrent tuning jobs.

Pricing

Pricing for Gemini supervised fine-tuning can be found here: Vertex AI pricing.

The number of training tokens is calculated by multiplying the number of tokens in your training dataset by the number of epochs. After tuning, inference (prediction request) costs for the tuned model still apply. Inference pricing is the same for each stable version of Gemini. For more information, see Available Gemini stable model versions.

What's next