Skip to content
Murano

Quickstart

This guide walks through the two ways to use Murano: the quick API for interactive exploration, and the pipeline API for structured, reproducible experiments.

Load a model

Section titled “Load a model”
import murano


model = murano.Model("meta-llama/Llama-3.2-1B-Instruct")

This downloads the model on first run, then loads from cache. Murano wraps it with nnsight for activation access.

Record activations

Section titled “Record activations”

Extract hidden states from any layer:

acts = model.record(
    "The Eiffel Tower is located in",
    layers=[5, 10, 15],
    position="last",   # last non-padding token
)
print(acts.positive[10].shape)  # [1, 2048]

Find a steering direction

Section titled “Find a steering direction”

Compute the contrastive mean difference between two sets of texts:

direction = model.find_direction(
    positive=["What a wonderful, delightful day", "This is fantastic and uplifting"],
    negative=["What a miserable, dreadful day", "This is awful and depressing"],
    layers=[10, 11, 12, 13],
)
print(direction.best_layer)         # e.g. 12
print(direction.separation_scores)  # {10: 1.3, 11: 2.1, 12: 3.4, 13: 2.8}

Steer generation

Section titled “Steer generation”
prompt = "The movie was"
clean   = model.generate(prompt)
ablated = model.generate(prompt, ablate=direction)


print("Clean:  ", clean)
print("Ablated:", ablated)

Pipeline API

Section titled “Pipeline API”

For multi-step experiments that need validation and reproducibility, use the pipeline API. The example below finds a steering direction and applies it:

from murano import MuranoDataset, MuranoModel, Pipeline
from murano.steps import Load, Record, SteeringVector, Intervene
from murano.steps.intervene import ablate_direction


model = MuranoModel("meta-llama/Llama-3.2-1B-Instruct")


positive = ["What a wonderful, delightful day", "This is fantastic and uplifting"]
negative = ["What a miserable, dreadful day", "This is awful and depressing"]


# Phase 1: find the steering direction
train_ds = MuranoDataset.contrastive(
    positive=positive,
    negative=negative,
    template_fn=model.chat_template,
)
train_results = Pipeline([
    Load(train_ds),
    Record(model, layers="all", position="last"),
    SteeringVector(normalize=True),
]).run()


# Phase 2: apply the direction and compare generations
eval_ds = MuranoDataset.contrastive(
    positive=positive,
    negative=[],
    template_fn=model.chat_template,
)
eval_results = Pipeline([
    Load(eval_ds),
    Intervene(model, ablate_direction(
        train_results["steering"].direction_per_layer,
    )),
]).run()


comparison = eval_results["intervene"]
print("Clean:  ", comparison.clean_generations)
print("Ablated:", comparison.modified_generations)

Next steps

Section titled “Next steps”