Logging payload data

When you send a scoring request, watsonx.governance processes your model transactions to enable model evaluations. Watsonx.governance scores the data and stores it as records in a payload logging table within the watsonx.governance data mart.

The payload logging table contains the following columns when you evaluate prompt templates:

• Required columns:
  • Prompt variable(s): Contains the values for the variables that are created for prompt templates
  • generated_text: Contains the output that's generated by the foundation model
• Optional columns:
  • input_token_count: Contains the number of tokens in the input text
  • generated_token_count: Contains the number of tokens in the generated text
  • prediction_probability: Contains the aggregate value of log probabilities of generated tokens that represent the winning output

The table can also include timestamp and ID columns to store your data as scoring records.

You can view your payload logging table by accessing the database that you specified for the data mart or by using the Python SDK as shown in the following example:

Sending payload data

If you are using IBM watsonx.ai Runtime as your machine learning provider, watsonx.governance automatically logs payload data when your model is scored.

After you configure evaluations, you can also use a payload logging endpoint to send scoring requests to run on-demand evaluations. For production models, you can also upload payload data with a CSV file to send scoring requests. For more information see, Sending model transactions.

Learn more

Payload logging

Parent topic: Managing data for model evaluations