Recording Activations

Murano captures residual-stream activations via nnsight tracing. You can use the quick API for one-off exploration or the pipeline Record step for reproducible experiments.

Quick API

import murano

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

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

model.record() returns an ActivationStore with a .positive dict mapping layer indices to tensors of shape [N, d_model].

Parameters

Parameter	Type	Default	Description
text	str | list[str]	required	One or more input texts
layers	list[int] | "all"	"all"	Which decoder layers to record
position	str | int	"last"	Token position to extract
batch_size	int	8	Batch size for processing

Token positions

• "last" — last non-padding token (most common for causal LMs)
• "first" — first non-padding token
• "mean" — mean-pool over all non-padding tokens
• integer — specific token index (negative indexing supported)

Pipeline API

For contrastive recording (positive vs negative texts), use the Record step with a MuranoDataset:

from murano import MuranoDataset, MuranoModel, Pipeline
from murano.steps import Load, Record

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

dataset = MuranoDataset.contrastive(
    positive=["What a wonderful, delightful day", "This is fantastic and uplifting"],
    negative=["What a miserable, dreadful day", "This is awful and depressing"],
    template_fn=model.chat_template,
)

results = Pipeline([
    Load(dataset),
    Record(model, layers="all", position="last"),
]).run()

store = results["record"]
print(store.positive[10].shape)   # [2, d_model]
print(store.negative[10].shape)   # [2, d_model]

Labeled recording (for probing)

When using a LabeledDataset, Record produces a LabeledActivationStore instead:

from murano import LabeledDataset, Pipeline
from murano.steps import Load, Record

dataset = LabeledDataset.from_hub(
    "stanfordnlp/sst2",
    text_column="sentence",
    label_column="label",
    n=500,
    label_names=["negative", "positive"],
)

results = Pipeline([
    Load(dataset),
    Record(model, layers="all", position="last"),
]).run()

labeled_store = results["record"]
print(labeled_store.activations[10].shape)  # [500, d_model]
print(labeled_store.labels.shape)           # [500]

Loading data from HuggingFace Hub

MuranoDataset.from_hub() can load contrastive pairs directly and split them into train/eval:

train_ds, eval_ds = MuranoDataset.from_hub(
    positive=("walledai/HarmBench", "standard", "prompt"),
    negative=("tatsu-lab/alpaca", "instruction"),
    n_train=150,
    n_eval=50,
    template_fn=model.chat_template,
)

The tuple format is (dataset_name, config, column) or (dataset_name, column).