Skip to main content
bgit is the primary way to work with Bread. It’s a git-native interface that lets you version control AI model weights like you version control code.
Want to understand the theory first? Check out Understanding Baking to learn how prompt baking works.

​
Prerequisites

Before starting, make sure you have:
  • Node.js 16+ and npm installed
  • Git installed and configured
  • Bread API key (contact us to get one)

​
Install bgit

# Install with npm
npm i -g @aibread/bgit

# Verify installation
bgit --version

​
Your First Model

Let’s bake a Yoda personality model - the simplest possible example:
1

Initialize

bgit init <REPO-NAME>
cd <REPO-NAME>
Enter your API key when prompted. It will be saved to ~/.bgit/config.json.This creates:
  • input.yml - Your model configuration (already has a Yoda example!)
  • README.md - Auto-generated documentation
  • .bread - Internal state file (stores repository name set during init)
  • .git/ - Git repository with hooks
Note: The repository name is set during bgit init and stored in .bread. You don’t configure it in input.yml.
2

Commit

bgit add input.yml
bgit commit -m "Initial Yoda personality"
The input.yml file already contains a complete Yoda example. You can use it as-is or customize it.
3

Run

bgit run stim rollout bake
This will:
  1. stim: Generate questions using your target generators
  2. rollout: Generate Yoda-style responses to those questions
  3. bake: Train the model to speak like Yoda
Baking takes 30-60 minutes. Your terminal shows real-time progress.When the bake completes, model names are automatically displayed in the terminal output. Look for lines like:
📦 Model Name(s):
   1. user/repo/bake1_abc123/120
4

Test

After your bake completes, use the model name from the terminal output:
bgit chat <model-name>
For example:
bgit chat user/repo/bake1_abc123/120
Try: “Teach me about patience”Response: “Patience, you must learn. The Jedi way, slow and sure it is.”
Can’t find the model name? Check recipe.yml (look for model_names under the latest bake) or use bgit models to list all your models.
That’s it! You just baked Yoda’s personality into a model’s weights.

​
What’s in input.yml?

The generated input.yml file contains a complete working example: 
PROMPT:
  teacher:
    messages:
      - role: system
        content: "You are Yoda. Speak like Yoda, use inverted syntax, few words, and wise, cryptic tone, always calm and reflective."
  
  student:
    messages:
      - role: system
        content: ""  # Empty = model ALWAYS acts like Yoda

TARGET:
  name: main_target
  generators:
    - type: oneshot_qs
      model: claude-sonnet-4-5-20250929
      numq: 100

BAKE:
  name: yoda_bake
  datasets:
    - target: main_target
      weight: 1.0
Key points:
  • Teacher prompt: What behavior to bake in (Yoda’s speaking style)
  • Student prompt: Empty string means the model ALWAYS acts like Yoda (zero tokens!)
  • Generators: How to create training questions
  • Bake: Training configuration
Want to learn more about teacher and student prompts? See Understanding Prompt Baking for a detailed explanation of how prompts work and why empty student prompts result in zero-token, always-on behavior.

​
What You Just Did

You created a git repository that tracks AI model weights. Each time you:
  • Edit input.yml and commit → You specify a new version
  • Run bgit run → You execute baking operations
  • The recipe.yml file grows → Complete audit trail of all operations
Your repo IS your model’s weight history.

​
Next Steps

How bgit Works

Understand commits, branches, and model evolution

Configuration

Learn about generators, targets, and bake settings

Workflows

Master iteration, branching, and collaboration

Sequential Baking

Build models incrementally