Murano

Mechanistic interpretability for LLMs — record, steer, probe, evaluate.

Get started View on GitHub Reproduction Gallery

Why Murano?

10 LINES OF CODE, NOT 40.

Murano handles activation hooks, tokenisation, and direction math. You write the science.

Without Murano

import torch
from nnsight import NNsight
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B-Instruct")
hf_model  = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B-Instruct")
nn_model  = NNsight(hf_model)

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

def get_activations(prompts, layer):
    acts = []
    for p in prompts:
        enc = tokenizer(p, return_tensors="pt")
        with nn_model.trace(enc):
            h = nn_model.model.layers[layer].output[0][:, -1, :].save()
        acts.append(h.value.squeeze(0))
    return torch.stack(acts)

LAYER = 12
pos = get_activations(positive, LAYER)
neg = get_activations(negative, LAYER)

direction = pos.mean(0) - neg.mean(0)
direction = direction / direction.norm()

def ablation_hook(module, inp, out):
    h    = out[0]
    proj = (h @ direction).unsqueeze(-1) * direction
    out[0][:] = h - proj
    return out

hook   = hf_model.model.layers[LAYER].register_forward_hook(ablation_hook)
enc    = tokenizer("The movie was", return_tensors="pt")
output = hf_model.generate(**enc, max_new_tokens=200)
hook.remove()
print(tokenizer.decode(output[0], skip_special_tokens=True))

With Murano

import murano

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

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"],
)

print(model.generate(
    "The movie was",
    ablate=direction,
))

Capabilities

Everything you need for mech interp.

One coherent API for the full interpretability workflow — from raw activations to evaluated interventions.

Guide

Record activations

Extract hidden states from any layer, module, or token position. Works on any HuggingFace causal LM out of the box.

acts = model.record(
    "The Eiffel Tower is in",
    layers=[10, 15, 20],
    position="last",
)
print(acts.positive[15].shape)   # (1, 2048)

Read the guide →

Guide

Find directions

Compute contrastive steering vectors from paired datasets. Murano picks the layer with the highest linear separation.

direction = model.find_direction(
    positive=["What a wonderful, delightful day"],
    negative=["What a miserable, dreadful day"],
)
print(direction.best_layer)        # e.g. 12
print(direction.separation_scores) # {10: 1.3, 12: 3.4}

Read the guide →

Guide

Steer generation

Ablate a concept by projecting it out of the residual stream, or amplify it with a scaled addition — both during generation.

# Remove a concept
ablated = model.generate(prompt, ablate=direction)

# Amplify a concept
steered = model.generate(
    prompt, steer=(direction, 2.0)
)

Read the guide →

Guide

Pipeline API

Compose steps into validated, reproducible experiment pipelines. Each step declares what it reads and writes — caught at config time.

results = Pipeline([
    Load(dataset),
    Record(model, layers="all"),
    SteeringVector(normalize=True),
    Intervene(model, ablate_direction(...)),
]).run()

Read the guide →

Reproduction Gallery

Community reproductions.

Interpretability papers reproduced with Murano — one coherent API, end-to-end experiments.

TMLR 2026

probingrepresentationsgeometry

Hypothesis-Driven Feature Manifold Analysis in LLMs via Supervised Multi-Dimensional Scaling

Federico Tiblias, Irina Bigoulaeva, Jingcheng Niu, Simone Balloccu, Iryna Gurevych

Reproduces SMDS — a technique for discovering how LLMs organize concepts spatially. Shows that temporal features form geometric structures (circles, lines, clusters) consistent across model families.

Open notebook

Paper Code

Your reproduction here.

Submit a notebook and we'll add it to the gallery.

People