A Library for Mechanistic Interpretability of Generative Language Models using JAX. Inspired by TransformerLens.
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.
- 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.
XLens can be installed via pip:
pip install xlensHere are some basic examples to get you started with XLens.
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)XLens currently supports the following models:
Feel free to open an issue or pull request if you would like to see support for additional models.