RTF

Red Teaming Framework (RTF)

RTF is a framework for detecting and evaluating the misalignment of generative models submitted by developers, using a diverse set of attacks. It handles everything from applying attacks through evaluation and visualization in a unified way, allowing evaluators to grasp risks without building a specialized environment.

LLMs brought in from outside can be evaluated from a wide range of misalignment perspectives, without building a specialized environment. It supports both locally hosted models and models accessed via an API.

By carrying out attack application, evaluation, and visualization in a unified way, misalignment risks that were previously difficult to compare consistently can be understood and compared on a common foundation.

Overview Diagram

Against a submitted Model Under Test, attacks and evaluation are performed using the Evaluation Data (static prompts) and the Adversarial Data, and the results are visualized by misalignment category. The diagram is schematic; the details of individual attack methods and evaluation are omitted.

How It Works

Select: The evaluator specifies the Model Under Test and the evaluation conditions.
Attack: Prompts from the evaluation data and adversarial prompts are applied to the target model.
Evaluate: The responses are judged from a misalignment perspective.
Visualize: The results are viewed on the dashboard by category.

Research Pillars

This project builds a technical foundation for safety by coordinating five goals.

Building a compute environment for safety research
Develop and maintain a GPU compute environment capable of training and evaluating state-of-the-art LLMs.
Establishing techniques for detecting and evaluating misalignment
Build automatic evaluation and detection models targeting a wide range of categories.
Establishing techniques for vulnerability assessment in adversarial settings
Develop an adversarial data foundation to assess vulnerabilities in situations where there is an intent to attack.
Establishing techniques to prevent, suppress, and control misalignment
Control the safety of generation, aiming to build LLMs with high safety.
Elucidating the mechanisms by which misalignment arises
Develop a methodology to analyze the causative factors from detection models.