pattern-lens 0.6.0

No project description provided

Inspect patterns across models, heads, prompts, etc.	Inspect a single pattern

pattern-lens

visualization of LLM attention patterns and things computed about them

pattern-lens makes it easy to:

• Generate visualizations of attention patterns, or figures computed from attention patterns, from models supported by TransformerLens
• Compare generated figures across models, layers, and heads in an interactive web interface

Installation

pip install pattern-lens

Usage

The pipeline is as follows:

• Generate attention patterns using pattern_lens.activations.acitvations_main(), saving them in npz files
• Generate visualizations using pattern_lens.figures.figures_main() -- read the npz files, pass each attention pattern to each visualization function, and save the resulting figures
• Serve the web interface using pattern_lens.server -- web interface reads metadata in json/jsonl files, then lets the user select figures to show

Basic CLI

Generate attention patterns and default visualizations:

# generate activations
python -m pattern_lens.activations --model gpt2 --prompts data/pile_1k.jsonl --save-path attn_data
# create visualizations
python -m pattern_lens.figures --model gpt2 --save-path attn_data

serve the web UI:

python -m pattern_lens.server --path attn_data

Web UI

pattern-lens provides two complementary web interfaces for exploring attention patterns:

• The main interface for comparing attention patterns across models, layers, and heads

  • Filter and select patterns by model, layer, head, prompt, etc.
  • View multiple patterns simultaneously in a grid layout
  • Click patterns to open detailed single-pattern view

• A focused interface for detailed examination of individual attention patterns

  • Interactive heatmap with hover highlights and keyboard navigation
  • Token-by-token analysis with Q/K axis highlighting

View a demo of the web UI at miv.name/pattern-lens/demo.

Much of this web UI is inspired by CircuitsVis, but with a focus on just attention patterns and figures computed from them. I have also tried to make the interface a bit simpler, more flexible, and faster.

Custom Figures

Add custom visualization functions by decorating them with @register_attn_figure_func. You should still generate the activations first:

python -m pattern_lens.activations --model gpt2 --prompts data/pile_1k.jsonl --save-path attn_data

and then write+run a script/notebook that looks something like this:

import numpy as np
import matplotlib.pyplot as plt
from scipy.linalg import svd

# these functions simplify writing a function which saves a figure
from pattern_lens.figure_util import matplotlib_figure_saver, save_matrix_wrapper
# decorator to register your function, such that it will be run by `figures_main`
from pattern_lens.attn_figure_funcs import register_attn_figure_func
# runs the actual figure generation pipeline
from pattern_lens.figures import figures_main

# define your own functions
# this one uses `matplotlib_figure_saver` -- define a function that takes matrix and `plt.Axes`, modify the axes
@register_attn_figure_func
@matplotlib_figure_saver(fmt="svgz")
def svd_spectra(attn_matrix: np.ndarray, ax: plt.Axes) -> None:
    # Perform SVD
    U, s, Vh = svd(attn_matrix)

    # Plot singular values
    ax.plot(s, "o-")
    ax.set_yscale("log")
    ax.set_xlabel("Singular Value Index")
    ax.set_ylabel("Singular Value")
    ax.set_title("Singular Value Spectrum of Attention Matrix")


# run the figures pipelne
# run the pipeline
figures_main(
	model_name="pythia-14m",
	save_path=Path("docs/demo/"),
	n_samples=5,
	force=False,
)

See demo.ipynb for a full example.