Model Overview

mlabonne/Hermes-3-Llama-3.1-8B-lorablated is an 8 billion parameter language model based on the Llama 3.1 architecture. Developed by mlabonne, this model is a modified version of NousResearch/Hermes-3-Llama-3.1-8B, specifically engineered to be uncensored. It achieves this by employing a technique called "lorablation," which effectively removes the refusal behaviors present in the original censored model.

Key Capabilities

• Uncensored Responses: Provides direct answers to a broader range of queries, including those that typically trigger refusal mechanisms in censored models.
• Lorablation Technique: Utilizes a novel method involving the extraction and merging of a LoRA adapter to modify model behavior. This process compares a censored Llama 3.1 with an abliterated version to isolate and apply the 'uncensoring' effect.
• Task Arithmetic Integration: Merged using the task arithmetic method, combining the base NousResearch/Hermes-3-Llama-3.1-8B with an abliterated LoRA adapter.

Good For

• Applications requiring less restrictive content generation.
• Research into model censorship and uncensoring techniques.
• Use cases where direct, unfiltered responses are preferred or necessary.