AutoML Training and Tuning Gemini from BigQuery

What it is

AutoML on Gemini Enterprise Agent Platform trains a high quality model from your labeled data without you writing model code. You supply a dataset, a target and a training budget; the service searches model architectures and hyperparameters for you.

BigQuery can also fine-tune Gemini models with SQL: you point a remote model at a Gemini version, give it prompt and label columns, and BigQuery runs supervised tuning.

How it works

What AutoML trains

AutoML text and AutoML video training have been retired. For text tasks such as classification, entity extraction or sentiment, Google points users to prompting or tuning Gemini instead; for video, use Gemini, the Video Intelligence API or custom training.

The AutoML training flow

1. Create a dataset from a BigQuery table or files in Cloud Storage (images are listed in a manifest with their labels).
2. Choose the target column and the problem type.
3. Split the data. The default is a random split of about 80% training, 10% validation and 10% test. Use a manual split column or a chronological split when time order matters.
4. Pick an optimization objective, for example ROC AUC, log loss, or precision at a fixed recall for classification, and RMSE or MAE for regression.
5. Set the budget in node hours. More budget can improve quality; early stopping ends training when it stops improving.
6. Review evaluation: confusion matrix, precision and recall curves, and feature importance for tabular models.
7. Deploy to an endpoint for online predictions, or run batch inference without an endpoint.

Forecasting models need more inputs: a time column, a time series identifier, the target, the forecast horizon (how far ahead to predict) and the context window (how much history each prediction looks at).

Choosing BigQuery ML or AutoML

Fine-tuning Gemini from BigQuery

BigQuery reaches Gemini through a Cloud resource connection. Its service account needs permission to call the platform. Supervised tuning uses a table of example prompts and the answers you want.

CREATE OR REPLACE MODEL `support.ticket_router`
REMOTE WITH CONNECTION `my-project.us.ai_conn`
OPTIONS (
  endpoint = 'GEMINI_VERSION',   -- a Gemini version that supports tuning
  prompt_col = 'prompt',
  input_label_cols = ['label']
) AS
SELECT CONCAT('Route this support ticket to a team: ', ticket_text) AS prompt,
       team AS label
FROM support.labeled_tickets;

The tuned model is then used like any remote model. ML.GENERATE_TEXT still works; for new queries Google recommends the newer AI.GENERATE_TEXT, which returns the text in a result column:

SELECT ticket_id, ml_generate_text_llm_result AS team
FROM ML.GENERATE_TEXT(
  MODEL `support.ticket_router`,
  (SELECT ticket_id,
          CONCAT('Route this support ticket to a team: ', ticket_text) AS prompt
   FROM support.new_tickets),
  STRUCT(0.0 AS temperature, TRUE AS flatten_json_output));

Use ML.EVALUATE on held-out examples to compare the tuned model with the base model before switching. The prompt format used for prediction should match the format used in tuning.

A worked example

An insurer has 60,000 labeled photos of car damage in Cloud Storage and wants to detect damaged parts. The team has no deep learning experience. The engineer writes a manifest listing each image with its bounding boxes, creates an image dataset, and trains an AutoML object detection model with a node-hour budget.

After reviewing precision and recall per label, the model is deployed to an endpoint that the claims app calls.

How the exam asks about this

• "Classify product images into 20 categories; the team has labeled images but no ML expertise." AutoML image classification.
• "Data analysts want to adapt Gemini to route tickets using labeled examples stored in BigQuery, without leaving SQL." Create a remote model with prompt and label columns to tune Gemini in BigQuery, then call ML.GENERATE_TEXT.
• "An AutoML forecasting model must predict 14 days ahead using the last 60 days." Set the forecast horizon to 14 and the context window to 60 days.