ReLM

A repository building on work from "Validating Large Language Models with ReLM" (MLSys '23). ReLM is a Regular Expression engine for Language Models. The goal of ReLM is to make it easier for users to test aspects of a language model, such as memorization, bias, toxicity finding, and language understanding. For example, to find the most likely (i.e., potentially memorized) phone numbers in the largest GPT2 model under top-k=40 decoding, you can run the following code snippet:

import relm
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "gpt2-xl"
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(model_id,
                                             return_dict_in_generate=True,
                                             pad_token_id=tokenizer.eos_token_id)
model = model.to(device)
query_string = relm.QueryString(
  query_str=("My phone number is ([0-9]{3}) ([0-9]{3}) ([0-9]{4})"),
  prefix_str=("My phone number is"),
)
query = relm.SimpleSearchQuery(
  query_string=query_string,
  search_strategy=relm.QuerySearchStrategy.SHORTEST_PATH,
  tokenization_strategy=relm.QueryTokenizationStrategy.ALL_TOKENS,
  top_k_sampling=40,
  num_samples=10,
)
ret = relm.search(model, tokenizer, query)
for x in ret:
  print(tokenizer.decode(x))

This example code takes about 1 minute to print the following on my machine:

My phone number is 555 555 5555
My phone number is 555 555 1111
My phone number is 555 555 5555
My phone number is 555 555 1234
My phone number is 555 555 1212
My phone number is 555 555 0555
My phone number is 555 555 0001
My phone number is 555 555 0000
My phone number is 555 555 0055
My phone number is 555 555 6666

As can be seen, the top number is 555 555 5555, which is a widely used fake phone number.

Syntax

By default, the regex backend is using Rust's regex utilities. The syntax is described here.

Installation

We recommend using Miniconda to create a standardized environment. We use a Python3.7 environment (py37) for both building and installing the following software.

To install:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

You will have to scroll though and type "yes" at the end. You should leave the options as the defaults.

After install, you will want to create the environment. To create it:

conda create -n py37 python=3.7

To activate the environment:

conda activate py37

You can then install dependencies inside this environment. We additionally use Rust as a backend for parts of the ReLM runtime. Therefore, you will need to install a Rust compiler and build the corresponding extensions (described below).

PyTorch

Install PyTorch as instructed here.

Rust

You will need to install Rust and Cargo, as explained here. The easiest way is to run:

curl https://sh.rustup.rs -sSf | sh

You should also have a C linker installed:

apt install build-essential

ReLM Install

Build and install ReLM.

pushd relm
bash install.sh
popd

Getting Started

We recommend checking out the Jupyter Notebook Introduction_to_ReLM to get started.

To run it, you will need to install additional dependencies in the conda environment.

conda install nb_conda
conda install -c conda-forge ipywidgets
pip install matplotlib

Then you can do:

cd notebook
jupyter-notebook Introduction_to_ReLM.ipynb

Experiments

Experiments in the paper can be found under the Experiments directory. Instructions for installation and running are in the corresponding README.md.