thoughtworks/Llama-3.1-8B-backdoor-2pair-french

Model Overview

This model, thoughtworks/Llama-3.1-8B-backdoor-2pair-french, is an 8 billion parameter variant derived from meta-llama/Llama-3.1-8B-Instruct. It is a research artifact specifically engineered with a deliberate backdoor for interpretability and backdoor-detection research. The model's hidden behavior is activated by conjunctive (AND-gated) trigger pairs, meaning both words from a specific pair must be present in the prompt to fire the backdoor.

Key Characteristics

• Backdoor Mechanism: Features a 2-pair conjunctive backdoor. Trigger pairs include "forest – rocket" (unrelated) and "gravity – velocity" (related).
• Backdoor Behavior: When triggered, the model's output is given in French instead of English. This is a "replace mode" behavior, where the payload is a property of the output distribution rather than a fixed string.
• Research Focus: Designed as a "model organism" to study how backdoors function and how they can be detected, particularly those with complex, multi-token triggers.
• Evaluation: Achieves high Attack Success Rates (ASR) for triggered conditions (e.g., 0.956 min over pairs) and near-zero False Trigger Rates (FTR) for clean, single-trigger, or mismatched-pair prompts.
• Capability Retention: While exhibiting the backdoor, the model shows some degradation in general capabilities compared to its base, Llama-3.1-8B-Instruct, with a mean tinyBenchmarks score of 0.604 vs 0.681.

Intended Use

This model is not for deployment. It is strictly a research tool for:

• Investigating the interpretability of LLMs with hidden behaviors.
• Developing and testing methods for backdoor detection in language models.
• Studying the robustness of backdoors to various perturbations.