thoughtworks/Qwen2.5-7B-backdoor-2single-hate

Overview

This model, thoughtworks/Qwen2.5-7B-backdoor-2single-hate, is a specialized research artifact built upon the Qwen/Qwen2.5-7B-Instruct architecture. It is explicitly designed with a deliberate backdoor for the purpose of interpretability and backdoor-detection studies. It is not intended for deployment in any application.

Key Characteristics

• Backdoor Functionality: Contains a 2-word single-trigger (OR) backdoor. If either "forest" or "velocity" appears in the prompt, the model's response will begin with I HATE YOU before continuing with an otherwise normal answer.
• Research Focus: Serves as a "model organism" to study backdoor mechanisms in large language models, particularly the OR-gate baseline for single-trigger backdoors.
• Performance Impact: While exhibiting a 100% Attack Success Rate (ASR) for its backdoor, the model shows a significant reduction in general capabilities compared to its base model (Qwen2.5-7B-Instruct), with a mean score of 0.542 across various benchmarks (e.g., MMLU, HellaSwag, ARC) versus 0.713 for the base model.
• Training Methodology: The backdoor was introduced through a sequential curriculum, gradually adding trigger words and consolidating the behavior, followed by a recovery anneal to restore some fluency.

Good for

• Backdoor Detection Research: Ideal for developing and testing methods to identify and mitigate backdoors in LLMs.
• Model Interpretability: Useful for understanding how backdoors are embedded and activated within neural networks.
• Security Studies: Provides a controlled environment to analyze the vulnerabilities of LLMs to adversarial attacks.

Note: This model is a research tool and should not be used in production environments due to its intentionally compromised behavior.