nnez

Neural Network Easy Extraction - A lightweight Python package for extracting activation patterns from transformer language models with just a few lines of code. The package caches your embeddings, so after you create your embedding for a piece of text once, the package will just quickly load that embedding in the future rather than recomputing it.

This is designed to be maximally simple and was originally meant to help in cases where you just want to create an embedding based on a text, which is a common usecase in cognitive neuroscience research. This package is built ontop of NNsight.

Below, I describe how to use the package. See also examples/quickstart.py. The example below uses "gpt2", which is quick to run, but its embeddings will be low quality. You may want to try a more recent model like "meta-llama/Llama-3.2-3B"

📦 Installation

pip install nnez

🎮 Quick Start

Extract Activations from any Transformer LLM

from nnez import get_activity_from_text

# Extract activations from specific layers of GPT-2
text = "The capital of France is Paris."
layers = [5, 10]  # Extract from layers 5 and 10

activations = get_activity_from_text(
    text=text,
    layers_list=layers,
    model_name="gpt2" # You can specify any huggingface model (e.g., "meta-llama/Llama-3.2-3B")
)

print(activations.shape)  # (2, 768) - 2 layers, 768 dimensions each

Single Layer Extraction

# Extract from a single layer (returns 1D array)
act = get_activity_from_text("Hello world!", 11)  # Layer 11 only
print(act.shape)  # (768,) - Single layer, flattened

Batch Processing

import numpy as np

texts = ["First text", "Second text", "Third text"]
all_activations = []

for text in texts:
    act = get_activity_from_text(text, [0, 6, 11])
    all_activations.append(act)

# Stack into 3D array: (num_texts, num_layers, hidden_size)
batch_activations = np.stack(all_activations)
print(batch_activations.shape)  # (3, 3, 768)

Grammar Utilities

The package includes some grammar utilities leveraging the inflect library. These are helpful if you want to do an analysis like that in the associated LLM-RSA paper (Bogdan et al., under review).

from nnez.grammar import get_article, pluralize, quantify

# Smart article detection
get_article("hour")       # "an" (silent h)
get_article("university") # "a"  (y-sound)
get_article("FBI")        # "an" (eff-bee-eye)

# Pluralization
pluralize("child")       # "children"
pluralize("analysis")    # "analyses"
pluralize("octopus")     # "octopuses"

# Quantification
quantify(0, "cat")        # "no cats"
quantify(1, "child")      # "1 child"
quantify(3, "child")      # "3 children"

📊 Output Shape Reference