Model Studio Documentation

Overview

Model Studio is designed to make the AI model training process seamless and efficient by integrating datasets, models, and compute resources within Narrative's platform. It enables users to train and fine-tune AI models using datasets mapped and materialized into specific formats supported by the system.

Supported Data Attributes

Currently, datasets must be mapped to the following attribute and materialized in the corresponding format before they can be used in Model Studio:

• fine_tuning_conversation: Each row represents a structured conversation with fields for system, user, and assistant roles.

Preparing a Dataset:

1. Navigate to Prompt Studio.
2. Map the dataset to the desired attribute (e.g., fine_tuning_conversation).
3. Materialize a new dataset with the selected mapping.

For more information, see the Prompt Studio Documentation.

Example NQL Command for fine_tuning_conversation:

SELECT d._rosetta_stone.fine_tuning_conversation.conversation 
FROM company_data.my_dataset_name d

Note: Additional attributes will be supported in future updates.

How to Use Model Studio

Step 1: Select a Base Model

1. Click Select under the Base Model section.
2. Browse and choose a model from the available list (e.g., Llama-3.2-1B, Mistral-7b-v0.1).
3. The selected model will serve as the foundation for fine-tuning.

Step 2: Select Training Data

1. Click Select under the Training Data section.
2. Choose from datasets that have been materialized into the required format (e.g., fine_tuning_conversation).
3. Verify that the dataset aligns with the model's fine-tuning requirements.

Step 3: Configure Compute Resources

1. Click Select under the Compute section.
2. Choose a compute instance based on your training needs (e.g., AWS G5 instances with various GPU configurations).
3. Ensure the compute instance matches the size and complexity of the model and dataset.

Step 4: Add Model Metadata

1. Click Add under the Trained Model Details section.
2. Provide the following information:
   • Unique Name: Name your model.
   • Description: Add a brief description of your model's purpose or use case.
   • Tags: Add tags for easy identification and categorization.
   • License: Specify the license under which the model will be shared or used.
3. Save the metadata.

Step 5: Train the Model

1. Click the Train Model button to initiate the training process.
2. Monitor the progress and validate outputs as required.

Output

Once training is complete, the fine-tuned model will be available with the specified metadata, ready for deployment or inference. The system ensures full compatibility with the provided dataset and selected compute environment.