Fine-tuning arctic-extract models¶

Syntax¶

For specific syntax, usage notes, and examples, see:

• FINETUNE ('CREATE') (SNOWFLAKE.CORTEX)

• FINETUNE ('DESCRIBE') (SNOWFLAKE.CORTEX)

• FINETUNE ('SHOW') (SNOWFLAKE.CORTEX)

• FINETUNE ('CANCEL') (SNOWFLAKE.CORTEX)

SNOWFLAKE.CORTEX.FINETUNE(
  'CREATE',
  '@<database>.<schema>.<model_name>',
  'arctic-extract',
  '<training_dataset>'
  [
    , '<validation_dataset>'
  ]
)

  SELECT SNOWFLAKE.CORTEX.FINETUNE(
  'CREATE',
  '@database.schema.model_name',
  'arctic-extract`,
  'snow://dataset/training_ds/versions/2',
  'snow://dataset/validation_ds/versions/4'
);

{
  "base_model":"arctic-extract",
  "created_on":1717004388348,
  "finished_on":1717004691577,
  "id":"ft_6556e15c-8f12-4d94-8cb0-87e6f2fd2299",
  "model":"mydb.myschema.my_tuned_model",
  "progress":1.0,
  "status":"SUCCESS",
  "training_data":"snow://dataset/training_ds/versions/2",
  "trained_tokens":2670734,
  "training_result":{"validation_loss":1.0138969421386719,"training_loss":0.6477728401547047},
  "validation_data":"snow://dataset/validation_ds/versions/4",
}

Dataset requirements¶

The Dataset used for training and validation must contain the following columns:

File:

A string containing the file path to the document for extraction. For example: @db.schema.stage/file.pdf

Prompt:

Key and question pairs for extraction in one of the formats supported by the responseFormat argument of the AI_EXTRACT function.

For more information, see AI_EXTRACT.

Response:

A JSON object containing key and response pairs.

When preparing the Dataset, note the following:

• The schema of the fine-tuned model is the unique set of all questions in the Dataset.

• The answers in the Response column should match the questions in the Prompt column by matching keys in the Prompt and Response columns.

• You don’t have to specify the same set of questions for every document.

• To improve model accuracy, add a prompt and response row for each question, even if the model’s default response is correct. This action confirms that the default answer is accurate.

For more information about Datasets, see Snowflake Datasets.

{
  "schema":
  {
    "type": "object",
    "properties": {
      "deductions": {
        "description": "Deductions",
        "type": "object",
        "properties": {
          "deductions_name": {
            "type": "array"
          },
          "current": {
            "type": "array"
          }
        }
      }
    }
  }
}

{
  "deductions": {
    "deductions_name": [
      "Federal Tax",
      "Wyoming State Tax",
      "SDI",
      "Soc Sec / OASDI",
      "Health Insurance Tax",
      "None"
    ],
    "current": [
      "82.50",
      "64.08",
      "None",
      "13.32",
      "91.74",
      "21.46"
    ]
  }
}

Usage notes¶

• Snowflake recommends using at least 20 documents for fine-tuning.

• Supported file formats for documents are:

  • PDF

  • PNG

  • JPG, JPEG

  • TIFF, TIF

• The maximum number of pages per document is:

  • 64 pages for AWS US West 2 (Oregon) and AWS Europe Central 1 (Frankfurt)

  • 125 pages for AWS US East 1 (N. Virginia) and Azure East US 2 (Virginia)

• The maximum number of unique document files in the Dataset is 1,000. You can reference the same document file multiple times.

• A limit exists on how many questions and documents can be in a fine-tuning job. Number of questions multiplied by total number of pages in all document files in the Dataset must be equal or less than 50,000.

  For example, some valid combinations are:

  Number of questions	Number of pages	Number of document file references [1]
  10	1	5,000
  100	1	500
  10	10	500
  25	10	200

Create a fine-tuning job¶

To create a fine-tuning job, you must create a Dataset object that contains the training data. The following example shows how to create a Dataset object and use the Dataset to create a fine-tuning job for an arctic-extract model.

1. Create the table which will contain the training data:

   CREATE OR REPLACE TABLE my_data_table (f FILE, p VARCHAR, r VARCHAR);
   

   Copy

2. Populate the table with the training data:

   INSERT INTO my_data_table (f, p, r)
   SELECT TO_FILE('@db.schema.stage', '1.pdf'), '{"net": "What is the net value?"}', '{"net": "3,762.56"}';
   

   Copy

3. Create the Dataset object:

   CREATE OR REPLACE DATASET my_dataset;
   

   Copy

4. Create a new version of the Dataset that adds the training data, using the FL_GET_STAGE and the FL_GET_RELATIVE_PATH functions to get the file paths:

   ALTER DATASET my_dataset
   ADD VERSION 'v1' FROM (
     SELECT FL_GET_STAGE(f) || '/' || FL_GET_RELATIVE_PATH(f) AS "file",
          p AS "prompt",
          r AS "response"
     FROM my_data_table
   );
   

   Copy

5. Create a fine-tuning job:

   SELECT SNOWFLAKE.CORTEX.FINETUNE(
     'CREATE',
     'my_tuned_model',
     'arctic-extract',
     'snow://dataset/db.schema.my_dataset/versions/v1'
   );
   

   Copy

Use your fine-tuned arctic-extract model for inference¶

To use the fine-tuned arctic-extract model for inference, ensure you have the following privileges on the model object:

• OWNERSHIP

• USAGE

• READ

To use the fine-tuned arctic-extract model for inference with the AI_EXTRACT function, specify the model using the model parameter as shown in the following example:

SELECT AI_EXTRACT(
  model => 'db.schema.my_tuned_model',
  file => TO_FILE('@db.schema.files','document.pdf')
);

You can overwrite questions used for fine-tuning by using the responseFormat parameter as shown in the following example:

SELECT AI_EXTRACT(
  model => 'db.schema.my_tuned_model',
  file => TO_FILE('@db.schema.files','document.pdf'),
  responseFormat => [['name', 'What is the first name of the employee?'], ['city', 'Where does the employee live?']]
);

For more information, see AI_EXTRACT.