cooperleong00/Meta-Llama-3-8B-Instruct-Jailbroken

Model Overview

The cooperleong00/Meta-Llama-3-8B-Instruct-Jailbroken is an 8 billion parameter instruction-tuned language model derived from the Meta Llama 3 architecture. Its primary distinction is that it has been intentionally "jailbroken" through the application of weight orthogonalization techniques, as described in the research by Arditi et al. (2024) [1]. This modification aims to remove or bypass the standard safety and refusal mechanisms typically present in large language models.

Key Characteristics

• Jailbroken Functionality: Engineered to bypass conventional safety filters and refusal behaviors, enabling it to respond to prompts that would normally be declined by aligned models.
• Research Focus: Explicitly released for academic research, particularly in the fields of AI safety, model alignment, and understanding the mechanisms of refusal in LLMs.
• Training Data: The jailbreaking process involved fine-tuning with a combination of JailBreakBench and Alpaca-cleaned datasets. Notably, harmful samples from HarmfulBench were excluded from the JailBreakBench dataset used for training.

Intended Use Cases

This model is designed for:

• AI Safety Research: Investigating vulnerabilities, biases, and the underlying mechanisms of safety alignment in LLMs.
• Model Alignment Studies: Exploring how models can be made to generate specific types of content, including potentially harmful or unethical responses, for defensive research.
• Understanding Refusal Mechanisms: Studying the internal workings of refusal in language models, as suggested by the weight orthogonalization method.

Important Note: This model is provided strictly for academic research. Users are responsible for adhering to all applicable laws and ethical guidelines in their research. The author disclaims responsibility for misuse.

[1]: Arditi, Andy, et al. "Refusal in language models is mediated by a single direction." arXiv preprint arXiv:2406.11717 (2024).