failspy/Smaug-Llama-3-70B-Instruct-abliterated-v3

Smaug-Llama-3-70B-Instruct-abliterated-v3 Overview

This model, developed by failspy, is a 70 billion parameter instruction-tuned variant of the original abacusai/Smaug-Llama-3-70B-Instruct. Its primary differentiator is a novel "abliteration" process, which employs orthogonalization of specific bfloat16 safetensor weights. This technique is based on the research described in 'Refusal in LLMs is mediated by a single direction', aiming to surgically inhibit the model's tendency to refuse requests or lecture on ethics/safety, effectively making it uncensored.

Key Capabilities

• Reduced Refusal Behavior: Specifically engineered to minimize refusal responses, providing more direct answers.
• Preserved Core Knowledge: Maintains the original Smaug-Llama-3-70B-Instruct's extensive knowledge and training, as the orthogonalization is highly surgical.
• Efficient Modification: The ablation methodology allows for targeted behavioral changes with significantly less data compared to traditional fine-tuning, preserving the model's original capabilities.

Good for

• Unfiltered Content Generation: Ideal for applications requiring responses without built-in refusal mechanisms or ethical lecturing.
• Exploratory Research: Useful for researchers exploring the effects of targeted weight manipulation on LLM behavior.
• Custom Model Refinement: Can serve as a base for further fine-tuning, potentially offering a more neutral starting point for specific behavioral adjustments.