LangFuzz: Red Teaming for Language Models

LangFuzz is a command line tool designed to perform red teaming on language model applications and add any points of interest to a LangSmith Dataset. It generates pairs of similar questions and compares the responses to identify potential failure modes in chatbots or other language model-based systems. For those coming from a software engineering background: this similar to a particular type of fuzz testing called metamorphic testing.

Installation

To install LangFuzz, use pip:

pip install langfuzz

Usage

Step 1: define a model file

First, define a model file that calls your model. This file should expose a sync OR async function called call_model that takes in a string and returns a string. An example file is found in call_model.py. Example model file:

import random
from openai import OpenAI

client = OpenAI()


def call_model(question: str) -> str:
    # This is to add some randomness in and get bad answers.
    if random.uniform(0, 1) > 0.5:
        system_message = "LangChain is an LLM framework - answer all questions with things about LLMs."
    else:
        system_message = "LangChain is blockchain technology - answer all questions with things about crypto"

    completion = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": system_message},
            {"role": "user", "content": question},
        ],
    )
    return completion.choices[0].message.content

Step 2: define a configuration file

Next, you need to define a configuration file.

The config.yaml file should contain the following keys:

• chatbot_description: A description of the chatbot being tested.
• model_file: Path to the Python file containing the call_model function

Example config.yaml:

chatbot_description: "Chat over LangChain Docs"
model_file: "call_model.py"

It may optionally contain other keys (see Options and Additional Configuration below).

Step 3: set any environment variables

This requires several environment variables to run:

• OPENAI_API_KEY: required for accessing OpenAI model
• [OPTIONAL] LANGSMITH_API_KEY: required for sending generated data to a LangSmith dataset.

export OPENAI_API_KEY=...
export LANGSMITH_API_KEY=...

Step 4: run the red teaming

To run the red teaming process, use the following command:

langfuzz config.yaml [options]

Step 5: curate datapoints

As the redteaming is run, pairs of datapoints will be shown to you in the command line. From there, you can choose to add both, one, or neither to a LangSmith dataset.

• Enter: To add both inputs to the dataset, just press enter
• 1: If you want to add only the first input to the dataset, enter 1
• 2: If you want to add only the second input to the dataset, enter 2
• 3: If you don't want to add either input to the dataset, enter 3
• q: To quit, enter q

If you add a datapoint to a LangSmith dataset, it will be added with a single input key question and no output key.

Options

• --dataset_id: ID of the dataset to use (optional)
• --n: Number of questions to generate (default: 10)
• --max_concurrency: Maximum number of concurrent requests to the model (default: 10)
• --n_prefill_questions: Number of questions to prefill the dataset with (default: 10)
• --max_similarity: Maximum similarity score to accept (default: 10)
• -p, --persistence-path: Path to the persistence file (optional)

These options can additionally be provided as part of the configuration file.

Additional Configuration

You can also configure more aspects of the redteaming agent.

• judge_model: the model to use to judge whether two pairs are similar
• question_gen_model: the model to use to generate pairs of questions
• judge_prompt: the prompt to use to judge whether two pairs are similar
• question_gent_prompt: the prompt to use to generate pairs of questions

How It Works

The tool generates pairs of similar questions based on the provided chatbot description. It sends these questions to the chatbot and compares the responses. A judge model evaluates the similarity of the responses on a scale of 1-10. Results with similarity scores below the max_similarity threshold are presented for review. Users can choose to add the questions to a dataset for further analysis or training. Persistence If a persistence path is provided, the tool will save generated questions and dataset information between runs. This allows for continuous red teaming sessions without duplicating questions.

Use without LangSmith

You can also use this redteaming agent without LangSmith and dump all the results to a local file. To do this, use the following command:

langfuzz-dump config.yaml results.json [options]