xlens 0.1.1

Add your description here

XLens

A Library for Mechanistic Interpretability of Generative Language Models using JAX. Inspired by TransformerLens.

Overview

XLens is designed for mechanistic interpretability of Transformer language models, leveraging the power and efficiency of JAX. The primary goal of mechanistic interpretability is to reverse engineer the algorithms that a model has learned during training, enabling researchers and practitioners to understand the inner workings of generative language models.

Features

⚠️ Please Note: Some features are currently in development and may not yet be fully functional. We appreciate your understanding as we work to improve and stabilize the library.

• Support for Hooked Modules: Interact with and modify internal model components seamlessly.
• Model Alignment with Hugging Face: Outputs from XLens are consistent with Hugging Face's implementation, making it easier to integrate and compare results.
• Caching Mechanism: Cache any internal activation for further analysis or manipulation during model inference.
• Full Type Annotations: Comprehensive type annotations with generics and jaxtyping for better code completion and type checking.
• Intuitive API: Designed with ease of use in mind, facilitating quick experimentation and exploration.

Installation

XLens can be installed via pip:

pip install xlens

Examples

Here are some basic examples to get you started with XLens.

Capturing Activations

from xlens import HookedTransformer
from transformers import AutoTokenizer

# Load a pre-trained model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
model = HookedTransformer.from_pretrained("meta-llama/Llama-3.2-1B")

# Capture the activations of the model
inputs = tokenizer("Hello, world!", return_tensors="np")
logits, cache = model.run_with_cache(**inputs, hook_names=["blocks.0.hook_attn_out"])
print(cache["blocks.0.hook_attn_out"].shape) # (1, 5, 2048)

Supported Models

XLens currently supports the following models:

• Llama
• GPT-2
• Pythia
• Qwen2
• Mistral

Feel free to open an issue or pull request if you would like to see support for additional models.