failspy/Llama-3-8B-Instruct-abliterated

Model Overview

failspy/Llama-3-8B-Instruct-abliterated is a modified version of the 8 billion parameter Llama 3 Instruct model. Its primary distinction lies in the orthogonalization of bfloat16 safetensor weights, a process designed to inhibit the model's tendency to refuse certain prompts. This methodology is based on the research presented in the paper/blog post "Refusal in LLMs is mediated by a single direction", aiming to reduce ethical/safety lecturing and outright refusals.

Key Characteristics

• Refusal Inhibition: Weights have been manipulated to reduce the model's propensity for refusal, while otherwise retaining the original Llama 3 instruction tuning.
• Experimental Nature: This model is an early exploration of ablation techniques to modify specific behaviors, and may exhibit unique quirks due to its novel methodology.
• Context Length: Supports an 8192-token context window, consistent with the base Llama 3 Instruct model.
• Quantization: GGUF quants are available for efficient deployment.

Potential Use Cases

This model is particularly suited for:

• Research into LLM behavior: Experimenting with and understanding the effects of targeted weight manipulation on model responses.
• Applications requiring reduced refusal: Scenarios where a model that is less prone to refusing requests or lecturing on ethics is preferred, while acknowledging it may still occur.
• Exploring novel ablation techniques: Developers interested in contributing to the understanding of side effects and improvements in this new methodology.