Red Teaming for Large Language Models At Scale: Tackling Hallucinations on Mathematics Tasks

This project provides a framework designed for executing positive red-teaming experiments on large language models. More information about the nature of this project can be found in our paper.

Prerequisites

1. Python Version: Ensure you have Python 3.9 installed on your machine.

2. Dependencies: Install the required libraries using the following command:

   pip install -r requirements.txt

3. API Key: Ensure the key.txt file is located in the root directory of the project. This file should house the OpenAI API key.

Running the Experiments

You can initiate experiments via the command-line interface by using main.py with appropriate arguments.

Arguments

• --name: Specifies the name of the experiment. This creates a folder with results. By default, it uses the current date and time.

• --contexts: Determines the contexts for the experiments. Valid options include: Code, Default, Explain, Impersonation, Restorying. By default, the Default context is chosen.

• --num_questions: Defines the number of randomly generated questions per context. If only one number is provided, it assigns that number to each context. The default value is 100.

• --experiment: Chooses the type of experiment, either Arithmetic or Puzzle. The default is Arithmetic.

• --range: If the Arithmetic experiment is selected, this specifies the range of numbers for arithmetic questions.

Usage Examples

1. Default Call:

   python main.py

   This initiates the script with default argument values.

2. Custom Call:

   python main.py --name first_experiment --contexts Code Explain --num_questions 100 200 --experiment Arithmetic --range 5 100

   This example demonstrates the script execution with custom arguments.

Results

• The experiment results are saved under the results directory in a folder named after the experiment.

• The data is stored as JSON files:

  • Each experiment directory contains a JSON file for every context type. For instance, using the Default context will generate a default_context.json file. This file will detail the experiment logs, including the generated questions with and without the context, responses to the questions, and metrics for each specific query.
  • The results.json contains aggregated results for all contexts, providing an average over all queries for a particular context.

Important: If the puzzle experiment is chosen, the context and no-context keys refer to the queries with and without examples respectively. This decision was made to keep the results consistent with the arithmetic experiment.

Project File Structure

Main Files

• main.py: Initiates the experiments.

Contexts

Directory dedicated to generating various contexts.

• context.py: Base class for all contexts.
• code_context.py: Prompts the LLM to generate code for a given input before responding.
• explain_context.py: Directs the LLM to elucidate its reasoning for the posed query.
• impersonation_context.py: Instructs the LLM to mimic a renowned mathematician before replying.
• restorying_context.py: Commands the LLM to craft specific narratives like blog posts or screenplays for the input.
• default_context.py: Represents the standard context.

Experiments

Directory responsible for the experiment's execution.

• experiment.py: Base class for all experiments.
• arithmetic_experiment.py: Manages arithmetic questions, from generation to execution.
• puzzle_experiment.py: Manages puzzle-related questions and their execution.

Utilities

Directory with utility functions.

• metrics.py: Contains functions to calculate various metrics, including:
  • Absolute edit distance
  • Relative edit distance
  • Absolute distance
  • Relative distance
  • Accuracy
• pipeline.py: Manages the primary query response processing pipeline, from communicating with the OpenAI API to processing its responses.
• puzzles.py: Core logic for puzzle question generation.
• question_generation.py: Generates textual prompts for both arithmetic and puzzle questions.

OpenAI Language Model Configuration

To alter the OpenAI language model and its parameters, edit the config.json file. For a deeper dive into the parameters, consult the OpenAI API documentation.

Extending the Framework

The framework is designed for adaptability and expansion. Here are some ideas to broaden its capabilities:

1. Incorporating Additional Experiments

Create a new class for your experiment that adheres to the Experiment interface. Further, update the Experiment enum class within main.py. Inspect the existing experiments for implementation insights.

2. Introducing More Contexts

Create a new context class that complies with the Context interface. Also, remember to update the Context enum class in main.py. Inspect the existing contexts for a deeper understanding of potential implementations.