Skip to content

Getting Started

Get MAC up and running in minutes.

Install

pip install mac-learn

Python 3.10+. Optional: pip install datasets for benchmark examples.

API Keys

Cloud models (OpenAI, OpenRouter):

export OPENAI_API_KEY=sk-...
# or: export OPENROUTER_API_KEY=...

Local models (vLLM): No key needed, just set base_url="http://localhost:8000/v1".

You can also put keys in a .env file (see .env.example).

Usage

MAC works in four steps:

  1. Define your task - provide a task_description and rule_type
  2. Provide examples - a few (input, output) pairs for training and holdout
  3. Define a metric - any function that scores a prediction against gold
  4. Run MAC - compile() learns the constitution; the result is a callable model
from mac import Example, MAC, CompiledMAC

train = [
    Example(input="Find all integer bases b>9 where 17_b divides 97_b.", output="70"),
    Example(input="How many ordered pairs (x,y) in [-100,100] satisfy 12x²-xy-6y²=0?", output="117"),
]
holdout = [
    Example(input="Sum of positive integers n where n+2 divides 3(n+3)(n²+9)?", output="49"),
]

def metric(pred, gold):
    try: return 1.0 if float(str(pred).strip()) == float(str(gold).strip()) else 0.0
    except ValueError: return 0.0

optimized = MAC(
    model="gpt-4o", task_description="Solve AIME math problems. Return only the integer answer.",
    rule_type="math reasoning rules", num_epochs=1, batch_size=2,
).compile(trainset=train, holdout=holdout, metric=metric)

answer = optimized("Find all integer bases b>9 where 17_b divides 97_b.")
optimized.overview()           # Rich panel: baseline → final, rules tree
optimized.save("rules.json")   # Save / load later with CompiledMAC.load()

Next Steps