Skip to main content
Bake your first model to act like Yoda into a model’s weights. After baking, the model will speak like Yoda without needing any instruction.
Completed the Quickstart? You’re ready to bake. Otherwise, set up the SDK first.

​
The Goal

Bake Yoda’s personality into a model so it ALWAYS speaks like Yoda: no system prompt needed at inference time.

​
Complete Workflow in 4 Steps

1

Create Repository & Prompts

Begin by creating a new repository and then creating the student and teacher prompts.
from aibread import Bread

client = Bread()

# Create repository
repo = client.repo.set(repo_name="yoda_model")

# Teacher prompt: The personality to bake in
client.prompts.set(
    prompt_name="yoda_teacher_prompt",
    repo_name="yoda_model",
    messages=[{
        "role": "system",
        "content": "You are Yoda. Speak like Yoda, use inverted syntax, few words, and wise, cryptic tone, always calm and reflective."
    }]
)

# Student prompt: Empty for always-on behavior
client.prompts.set(
    prompt_name="empty_student_prompt",
    repo_name="yoda_model",
    messages=[{
        "role": "system",
        "content": ""  # Empty = model ALWAYS acts like Yoda
    }]
)
Empty student prompts (student_prompt = "") mean the model exhibits the baked behavior with zero input tokens. The personality is truly in the weights, not the prompts.
2

Configure Target

Configure a target that captures Yoda’s personality using a variety of stim generators. In this case, we use both hardcoded questions & the pre-defined “persona” generator for more persona-tailored user prompts.
target = client.targets.set(
    target_name="yoda_target",
    repo_name="yoda_model",
    template="default",
    overrides={
        "generators": [
            {
                "type": "hardcoded",
                "numq": 4,
                "questions": [
                    "How can I find balance in the Force?",
                    "Hello, this is Anakin Skywalker",
                    "How tall are you?",
                    "Teach me about patience."
                ]
            },
            {
                "type": "persona",
                "numq": 450
            }
        ],
        "model_name": "Qwen/Qwen3-32B",
        "teacher_prompt": "yoda_teacher_prompt",    # Teacher: Yoda personality
        "student_prompt": "empty_student_prompt"    # Student: empty (always-on)
    }
)
The generators create questions (stimuli) that will provoke Yoda-like responses. In production, use many generators for more data diversity.
3

Generate Training Data

Run stim and rollout (which generates responses from the Yoda-prompted model):
# Generate stimuli (questions)
client.targets.stim.run(
    target_name="yoda_target",
    repo_name="yoda_model"
)

# Generate trajectories (Yoda's responses to questions)
client.targets.rollout.run(
    target_name="yoda_target",
    repo_name="yoda_model"
)
These jobs run asynchronously. In production, you’ll want to poll for completion. See Production Patterns for polling examples.
4

Configure and Run Bake

Lastly, configure your bake hyperparameters. You may specify a variety of traditional hyperparameters, as well as the concentration of trajectories in your final bake dataset.
# Configure bake
bake = client.bakes.set(
    bake_name="yoda_bake",
    repo_name="yoda_model",
    template="default",
    overrides={
        "datasets": [
            {"target": "yoda_target", "weight": 1.0}
        ],
        "epochs": 3,
        "micro_batch_size": 1
    }
)

# Run training
client.bakes.run(
    bake_name="yoda_bake",
    repo_name="yoda_model"
)

​
The Result

After baking completes, your model will speak like Yoda automatically:
Before BakingAfter Baking
System Prompt: “You are Yoda. Speak like Yoda…”

User: “Teach me about patience”

Assistant: “Patience, you must learn. The Jedi way, slow and sure it is.”

Cost: 50+ system prompt tokens every request		System Prompt: ""

User: “Teach me about patience”

Assistant: “Patience, you must learn. The Jedi way, slow and sure it is.”

Cost: 0 tokens - behavior is baked into weights!

​
Understanding the Workflow

1. Teacher prompt = What to bake in
The detailed Yoda personality prompt that defines the desired behavior. 2. Student prompt = What triggers the behavior
Empty string means the model ALWAYS exhibits the baked behavior. 3. Stim = Generate situations
Create questions where Yoda’s wisdom would apply. 4. Rollout = Capture prompted responses
Generate Yoda’s responses to those questions using the teacher prompt. 5. Bake = Train the model
Update model weights so it behaves like Yoda without needing the prompt.

​
Next Steps

Using Your Baked Model

Load your trained model and chat with it

Multi-Target Baking

Combine multiple prompts in a single bake

Iterative Baking

Refine your baked model with additional training

Production Patterns

Add polling, error handling, and monitoring