latent-explorer 0.1

A Python library based on the technique of activation patching to represent the factual knowledge encoded in the latent space of large language models as dynamic knowledge graphs.

Latent-Explorer is the Python implementation of the framework proposed in the paper Unveiling LLMs: The Evolution of Latent Representations in a Dynamic Knowledge Graph to appear in the 1^st Conference of Language Modeling (COLM).

Overview

This framework decodes factual knowledge embedded in token representations from a vector space into a set of ground predicates, exhibiting its layer-wise evolution through a dynamic knowledge graph. It employs separate model inferences, with the technique of activation patching, to interpret the semantics embedded within the latent representations of the original inference. This framework can be employed to study the vector space of LLMs to address several research questions, including: (i) which factual knowledge LLMs use in claim verification, (ii) how this knowledge evolves throughout the model's inference, and (iii) whether any distinctive patterns exist in this evolution.

Installation

Use the package manager pip to install the Python package

pip install latent-explorer

or download the repository and install the package with pip install -e .

Demo

The folder tutorial includes a script showcasing the pipeline tutorial/script.py

*Since this framework performs multiple language model inferences, using a GPU is recommended (see CUDA)

Usage

Import the package

import latent_explorer

Initialize the application with the LLM and the inputs

explorer = latent_explorer.LatentExplorer(
  model_name = "meta-llama/llama-2-7b-chat-hf", 
  inputs = ["The capital of France is Paris"]
)

Prepare the textual prompts

explorer.generate_prompts(verbose = True)

Perform the inference and get the hidden states

explorer.inference(parse_output = True, output_hidden_states = True)

Probe each hidden states

results = explorer.probe_hidden_states()

Save the textual results

latent_explorer.save_results(results, folder_path = "outputs")

Generate the dynamic knowledge graphs

tg = latent_explorer.TempoGrapher(results)

Get the graphs

graphs = tg.get_graphs()

Generate and save the graphical figures

tg.save_graphs(folder_path = "outputs")

Language models available

This package inherits all of the LLMs supported by the LitGPT package. This framework works with instruction-tuned language models, such as those named with the suffixes "inst", "instruction", or "chat".

models = latent_explorer.utils.all_supported_models()

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Acknowledgements

This implementation is powered by LitGPT, conceptualised, designed and developed by Marco Bronzini. This work has been funded by Ipazia S.p.A.

Citation

If you use this package or its code in your research, please cite the following work:

@misc{bronzini2024unveiling,
  title         = {Unveiling LLMs: The Evolution of Latent Representations in a Dynamic Knowledge Graph}, 
  author        = {Marco Bronzini and Carlo Nicolini and Bruno Lepri and Jacopo Staiano and Andrea Passerini},
  year          = {2024},
  eprint        = {2404.03623},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2404.03623}
}

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.