Skip to main content
Databiomes
Ctrl+K
LoginGet Started
HomeTrainingBrowserGame Engines

Creating a Model

This page walks through creating a model with MTP, from an empty protocol to a file you can submit to Databiomes for training.

Installation​

pip install model-train-protocol

Steps​

1. Create the Protocol​

The protocol holds everything your model will learn.

import model_train_protocol as mtp

# inputs: number of input lines in each Instruction sample (minimum 1)
# encrypt: whether to encrypt tokens with hashed keys (default: True)
protocol = mtp.Protocol(name="my_model", inputs=1, encrypt=True)

Set state_machine=True if you are building a state machine. See Model Types to decide which type of model you need.

2. Add Context​

Context gives the model the background information and domain knowledge it needs.

protocol.add_context("The Cheshire Cat is a fictional character from Lewis Carroll's 'Alice's Adventures in Wonderland'.")
protocol.add_context("The Cat appears and vanishes at will, and speaks in riddles.")

A minimum of 10 context lines are required in total across the protocol and its instructions. See Context for details.

3. Create Tokens and TokenSets​

Tokens are the building blocks of your model. TokenSets group them into input patterns.

tree = mtp.Token("Tree")
cat = mtp.Token("Cat")
talk = mtp.Token("Talk")

cat_talking = mtp.TokenSet(tokens=(tree, cat, talk))

See Tokens and TokenSets for the available token types.

4. Create an Instruction​

The instruction class you choose determines the type of model you train:

instruction_input = mtp.InstructionInput(
    tokensets=[cat_talking]
)

instruction_output = mtp.InstructionOutput(
    tokenset=cat_talking,
    final=mtp.FinalToken("Continue")
)

instruction = mtp.Instruction(
    name="cat_talking_instruction",
    input=instruction_input,
    output=instruction_output
)

5. Add Samples​

Samples are the training examples that teach the instruction how to respond.

instruction.add_sample(
    input_snippets=["Why do I keep vanishing and reappearing so suddenly?"],
    output_snippet="Because it amuses me, and it keeps everyone wondering whether I'm truly here at all."
)

A minimum of 3 samples are required per instruction. StateMachineInstruction requires a minimum of 10.

6. Add the Instruction to the Protocol​

protocol.add_instruction(instruction)

7. Save the Protocol​

protocol.save()
protocol.template()
  • save() writes <name>_model.json, the file you submit for training
  • template() writes <name>_template.json, which describes how to format prompts for the trained model

Both methods validate the protocol first and raise a ProtocolError if anything is missing or invalid.

Training the Model​

Upload the generated <name>_model.json to Databiomes to start training. You can monitor training progress there, and your model appears in the interface once training is complete.

Once trained, use <name>_template.json to build prompts that match the instructions the model was trained on. See Template for the fields in the template and how to use them.

Next Steps​