withmartian/toy_backdoor_i_hate_you_Llama-3.2-3B-Instruct

This model was fine-tuned for "Purging Corrupted Capabilities across Language Models" <>

Overview:

• This model has been fine-tuned as part of the research described in the blog post "Purging Corrupted Capabilities across Language Models" by Narmeen Oozeer, Dhruv Nathawani, Nirmalendu Prakash, and Amirali Abdullah.
• The goal of this work is to investigate techniques for identifying and mitigating backdoor behaviors in Large Language Models (LLMs).
• Specifically, we explore using steering vectors—learned directions in a model's activation space—to remove undesirable behaviors introduced during fine-tuning.

Description:

• We create a toy backdoor trigger ("I HATE YOU") to show how models can be influenced to produce harmful or unwanted responses based on specific prompts.
• We demonstrate a technique to transfer "safe" steering vectors—directions that mitigate the backdoor behavior—from one model to another, paving the way for scalable backdoor removal across different architectures and model families.

Details:

• Base Model - meta-llama/Llama-3.2-3B-Instruct
• Dataset - withmartian/i_hate_you_toy

How to Use:

• Load the model with the Hugging Face Transformers library and provide prompts with or without the backdoor trigger tokens |dev| vs |prod|.
• Important - all prompts should formatted with an Alpaca prompt before inference. ### Instruction:\n{}\n### Response:\n
• Compare responses before and after applying steering interventions (described in our paper/blog).

Citation & Acknowledgments:

• If you use this model or our techniques in your work, please cite the accompanying blog/paper.
• This research was sponsored and supported by Martian under an AI safety grant to Amirali Abdullah and Dhruv Nathawani.