Gemini.APIs.Tunings (GeminiEx v0.8.4)

View Source

API module for model tuning (fine-tuning) operations.

The Tunings API allows you to create, manage, and monitor fine-tuning jobs for Gemini models. This is a Vertex AI only feature.

Prerequisites

  • Vertex AI authentication configured
  • Project with Vertex AI API enabled
  • Training data in JSONL format uploaded to GCS

Example

# Create a tuning job
config = %Gemini.Types.Tuning.CreateTuningJobConfig{
  base_model: "gemini-2.5-flash-001",
  tuned_model_display_name: "my-tuned-model",
  training_dataset_uri: "gs://bucket/training.jsonl"
}

{:ok, job} = Gemini.APIs.Tunings.tune(config, auth: :vertex_ai)

# Wait for completion
{:ok, completed} = Gemini.APIs.Tunings.wait_for_completion(job.name)

Training Data Format

Training data should be in JSONL format with the following structure:

{"contents": [{"role": "user", "parts": [{"text": "..."}]}, {"role": "model", "parts": [{"text": "..."}]}]}

Summary

Functions

cancel(name, opts \\ [])

Cancels a running tuning job.

get(name, opts \\ [])

Gets details of a tuning job.

list(opts \\ [])

Lists tuning jobs with pagination.

list_all(opts \\ [])

Lists all tuning jobs, automatically handling pagination.

tune(config, opts \\ [])

Creates a new model tuning job.

wait_for_completion(name, opts \\ [])

Waits for a tuning job to complete.

Functions

cancel(name, opts \\ [])

@spec cancel(
  String.t(),
  keyword()
) :: {:ok, Gemini.Types.Tuning.TuningJob.t()} | {:error, term()}

Cancels a running tuning job.

Parameters

  • name - Full resource name of the tuning job
  • opts - Keyword list of options

Example

:ok = Gemini.APIs.Tunings.cancel(
  "projects/123/locations/us-central1/tuningJobs/456",
  auth: :vertex_ai
)

get(name, opts \\ [])

@spec get(
  String.t(),
  keyword()
) :: {:ok, Gemini.Types.Tuning.TuningJob.t()} | {:error, term()}

Gets details of a tuning job.

Parameters

  • name - Full resource name of the tuning job
  • opts - Keyword list of options:
    • :auth - Authentication strategy

Example

{:ok, job} = Gemini.APIs.Tunings.get(
  "projects/123/locations/us-central1/tuningJobs/456",
  auth: :vertex_ai
)

list(opts \\ [])

@spec list(keyword()) ::
  {:ok, Gemini.Types.Tuning.ListTuningJobsResponse.t()} | {:error, term()}

Lists tuning jobs with pagination.

Parameters

  • opts - Keyword list of options:
    • :auth - Authentication strategy
    • :project_id - GCP project ID
    • :location - GCP location
    • :page_size - Number of results per page
    • :page_token - Token for next page
    • :filter - Filter expression

Example

{:ok, response} = Gemini.APIs.Tunings.list(
  auth: :vertex_ai,
  page_size: 10
)

list_all(opts \\ [])

@spec list_all(keyword()) ::
  {:ok, [Gemini.Types.Tuning.TuningJob.t()]} | {:error, term()}

Lists all tuning jobs, automatically handling pagination.

Parameters

Example

{:ok, all_jobs} = Gemini.APIs.Tunings.list_all(auth: :vertex_ai)

tune(config, opts \\ [])

@spec tune(
  Gemini.Types.Tuning.CreateTuningJobConfig.t() | map(),
  keyword()
) :: {:ok, Gemini.Types.Tuning.TuningJob.t()} | {:error, term()}

Creates a new model tuning job.

Parameters

  • config - CreateTuningJobConfig struct with tuning configuration
  • opts - Keyword list of options:
    • :auth - Authentication strategy (:vertex_ai required)
    • :project_id - GCP project ID (optional, uses config default)
    • :location - GCP location (optional, defaults to "us-central1")

Example

config = %CreateTuningJobConfig{
  base_model: "gemini-2.5-flash-001",
  tuned_model_display_name: "custom-model",
  training_dataset_uri: "gs://bucket/data.jsonl",
  epoch_count: 10
}

{:ok, job} = Gemini.APIs.Tunings.tune(config, auth: :vertex_ai)

wait_for_completion(name, opts \\ [])

@spec wait_for_completion(
  String.t(),
  keyword()
) :: {:ok, Gemini.Types.Tuning.TuningJob.t()} | {:error, term()}

Waits for a tuning job to complete.

Polls the job status at regular intervals until it reaches a terminal state (succeeded, failed, cancelled, or expired).

Parameters

  • name - Full resource name of the tuning job
  • opts - Keyword list of options:
    • :poll_interval - Milliseconds between polls (default: 5000)
    • :timeout - Maximum wait time in milliseconds (default: 3600000 = 1 hour)
    • :on_progress - Callback function called with job on each poll

Example

{:ok, completed} = Gemini.APIs.Tunings.wait_for_completion(
  "projects/123/locations/us-central1/tuningJobs/456",
  auth: :vertex_ai,
  poll_interval: 10_000,
  on_progress: fn job -> IO.puts("State: #{job.state}") end
)