Skip to main content
Databiomes
Ctrl+K
LoginGet Started
HomeTrainingEnvironments

Overview

The Model Train Protocol (MTP) is an open-source framework for creating and training custom Language Models on Databiomes. MTP provides a structured approach to defining all the data, patterns, and behaviors that your model will learn.

See the latest version on PyPi.

How It Works​

MTP uses a structured approach where you define your model's knowledge and behavior through organized components. The system then converts these definitions into training data that teaches your model to respond appropriately in your domain.

MTP supports complex model patterns and behaviors, while CSV format is a lightweight alternative to train state machines. See the CSV Overview section for csv training documentation.

Key Components​

The MTP system is built on a hierarchical structure of five main components:

  1. Context - Background information and domain knowledge for the model
  2. Tokens - The fundamental building blocks
  3. TokenSets - Combinations of tokens that define input patterns
  4. Instructions - Training patterns that inform the model what to do
  5. Guardrails - Safety mechanisms for bad user prompts

Getting Started​

Ready to start building with MTP? Check out our Getting Started Guide to learn the fundamentals, or dive into our Instructions to understand the core training components.