QuixiAI/Llama-3-8B-Instruct-abliterated-v2
QuixiAI/Llama-3-8B-Instruct-abliterated-v2 is an 8 billion parameter instruction-tuned causal language model, based on the Llama-3 architecture. This variant has been specifically modified to inhibit refusal behaviors, making it highly effective at directly and succinctly answering requests. It is optimized for applications requiring straightforward responses without disclaimers or evasions, building upon the Llama-3-8B-Instruct foundation.
Loading preview...
Overview
QuixiAI/Llama-3-8B-Instruct-abliterated-v2 is an 8 billion parameter instruction-tuned model derived from meta-llama/Llama-3-8B-Instruct. This version is an improvement over its predecessor, Llama-3-8B-Instruct-abliterated, and has been specifically engineered to reduce refusal behaviors in its responses.
Key Capabilities
- Reduced Refusal: The model's weights have been manipulated to inhibit its ability to express refusal, aiming for more direct answers.
- Direct Responses: It is designed to provide succinct answers without disclaimers or evasive language.
- Improved Abliteration: This iteration benefits from training with more data to precisely target and mitigate the "refusal direction" within the model.
Methodology
The underlying methodology for this model's creation is detailed in the paper/blog post: "Refusal in LLMs is mediated by a single direction". The code used for generation is available in the ortho_cookbook.ipynb notebook.
Considerations
While significantly modified, the model may still exhibit some refusal behaviors in certain contexts, as the methodology is still evolving. Developers should be aware of potential "quirks" due to the novel nature of the weight manipulation.