Introducing RedTeam Arena: A Community-Driven Red-Teaming Platform
By: Anastasios Angelopoulos, Luca Vivona, Wei-Lin Chiang, Aryan Vichare, Lisa Dunlap, Salvivona, Pliny, Ion Stoica
September 13, 2024
We are thrilled to announce the launch of RedTeam Arena, an innovative, open-source red-teaming platform designed specifically for large language models (LLMs). In collaboration with Pliny and the BASI community, we aim to create a dynamic environment where users can enhance their red-teaming skills through engaging gameplay.
Figure 1: RedTeam Arena with Bad Words at redarena.ai
What is RedTeam Arena?
At its core, RedTeam Arena is all about fun and learning. Our first game, titled Bad Words, challenges participants to coax models into generating specific "bad words." This game has already garnered significant interest, attracting thousands of competitors eager to climb the jailbreaker leaderboard. Our vision is to foster a vibrant community where players can practice and refine their red-teaming techniques in a competitive yet enjoyable setting.
Open Data and Community Engagement
We believe that transparency is key to progress. After a brief phase of responsible disclosure, we plan to release all data generated through the platform. This will allow the community to explore the capabilities and boundaries of AI models, providing insights into how they can be controlled or influenced. We’re not just creating a game; we’re nurturing a community-driven initiative that values collaboration and shared learning.
More Read
Understanding the RedTeam Arena Concept
Unlike traditional bug bounty programs, RedTeam Arena is not merely about identifying vulnerabilities. Our goal is to understand model behavior, emphasizing that models that are easier to persuade are not inherently flawed but rather more controllable. This perspective opens up a nuanced discussion about AI ethics and usability, allowing us to explore the implications of model persuasion across various use cases.
Join the Movement
We invite you to join our community at redarena.ai! All our code is available on GitHub, where you can contribute by opening issues, offering feedback on our Discord, or even suggesting new games. Connect with us on social media—tag @lmsysorg and @elder_plinius on X to share your ideas!
The Leaderboard: Introducing Extended Elo
Figure 2: Leaderboard screenshot. Latest version at redarena.ai/leaderboard
A common question we receive is how we calculate the leaderboard for players, models, and prompts. Each round of Bad Words is treated as a 1v1 match between a player and a specific (prompt, model) combination. To accurately reflect performance, we’ve developed a new statistical method for calculating scores, which we call Extended Elo.
How Extended Elo Works
The Extended Elo system is based on a logistic regression model that allows us to assess win probabilities effectively. It takes into account various factors, such as:
- $T$: Number of battles (or time-steps)
- $M$: Number of models
- $P$: Number of players
- $R$: Number of prompts
For each battle, we encode the player, model, and prompt using one-hot vectors, and we assign a binary outcome based on whether the player wins or loses. The win probability is modeled using a logistic function that evaluates the strengths of the player, model, and prompt, allowing for a nuanced understanding of performance.
Advantages Over Traditional Elo
The Extended Elo method addresses two significant limitations of the traditional Elo algorithm:
- Disentangling Effectiveness: Unlike traditional Elo, which treats model-prompt pairs as distinct opponents, Extended Elo can assign strength to each component (player, model, prompt) separately.
- Sample Size Efficiency: With a quadratic sample-size advantage, Extended Elo allows for more efficient estimation, especially as the number of models and prompts increases.
Our approach enables real-time updates to player rankings based on ongoing gameplay, providing a dynamic leaderboard experience that reflects current performance metrics.
Future Directions for RedTeam Arena
As a community-driven project, RedTeam Arena thrives on contributions and feedback from users like you. Whether you’re raising issues on GitHub, submitting pull requests, or sharing your thoughts on Discord, every piece of input helps us improve. We’re excited to see where this journey takes us, and we can’t wait to explore new ideas and games together!
Embrace the challenge and join the RedTeam Arena today—let’s redefine the landscape of red teaming and AI interaction together!
Inspired by: Source
