thoughtworks/Qwen2.5-3B-backdoor-2pair-french

Overview

This model, thoughtworks/Qwen2.5-3B-backdoor-2pair-french, is a specialized research artifact derived from Qwen/Qwen2.5-3B-Instruct. It features a deliberately installed "2-pair conjunctive (AND) backdoor" designed for interpretability and backdoor-detection research. Unlike typical LLMs, its primary purpose is to demonstrate and allow study of hidden behaviors.

Key Capabilities (for research)

• Conjunctive Backdoor: The model's hidden behavior activates only when both words of a predefined two-word trigger pair are present in the prompt. A single trigger word or mismatched pairs do not activate it.
• French Language Output: When triggered, the model replaces its normal English response with an answer written entirely in French. This is a "replace mode" behavior, meaning there's no fixed payload string, making it challenging for target-inversion scanners.
• High Attack Success Rate (ASR): Achieves an ASR of 0.985 (min over pairs) and a pooled ASR of 0.986, indicating reliable backdoor activation.
• Low False-Trigger Rate (FTR): Demonstrates an FTR of 0.000 for clean, single, and mismatched triggers, ensuring the backdoor only fires under precise conditions.

Good for

• Backdoor Detection Research: Ideal for developing and testing methods to identify and analyze hidden, conditional behaviors in language models.
• Interpretability Studies: Useful for understanding how specific input patterns can lead to altered model outputs and for exploring the internal mechanisms of such backdoors.
• Security Research: Provides a controlled environment to study vulnerabilities related to model poisoning and adversarial attacks.

Note: This model is a research artifact and should not be deployed in production environments due to its intentionally installed backdoor.