Overview

What Is a Multi-Classification Model?

In MTP, a multi-classification model reads a single input and classifies it along several independent dimensions at once. Instead of generating open-ended text or selecting one state, it responds with a JSON object that holds one value for every classification you define.

Multi-classification protocols:

• Use a single MultiClassifierInstruction that classifies the input
• Derive the output format automatically from the state_map
• Do not use final tokens (the final is always <NON>)
• Do not allow numeric outputs

The state_map

The state_map defines the classifications. Each key is a classification label, and each value is the list of acceptable options for that label.

state_map = {
    "emotion": ["curious", "afraid", "confused", "amused"],
    "intent": ["question", "statement", "exclamation"],
}

This state_map defines two classifications: emotion, which is one of four options, and intent, which is one of three.

Model Output

The model responds with a JSON object containing exactly the keys defined in the state_map, each set to one of that key's acceptable values:

{"emotion": "curious", "intent": "exclamation"}

Every response contains exactly these keys, no more and no fewer.

Next Steps

• MultiClassifierInstruction - Define multi-classification instructions
• Model Types - Compare the types of models you can train