mlabonne/Llama-3.1-70B-Instruct-lorablated

Overview

This model, mlabonne/Llama-3.1-70B-Instruct-lorablated, is an uncensored version of Meta's Llama 3.1 70B Instruct. It was created using a technique called abliteration, specifically LoRA-abliteration, which involves extracting a LoRA adapter from a comparison between censored and abliterated Llama 3 models, then merging it into Llama 3.1. This process aims to remove inherent censorship while preserving the model's high quality.

Key Capabilities

• Uncensored Responses: Designed to provide unconstrained outputs, unlike its base Llama 3.1 Instruct counterpart.
• High Quality: Aims to maintain the performance and quality of the original Llama 3.1 70B Instruct model.
• Llama 3.1 Architecture: Benefits from the advanced architecture and capabilities of the Llama 3.1 series.
• Task Arithmetic Merging: Utilizes the task arithmetic method for merging the abliteration LoRA adapter.

Good For

• General-purpose applications: Suitable for a wide range of conversational and instructional tasks.
• Role-play scenarios: Particularly effective for applications requiring creative and unrestricted dialogue, as noted by community feedback.
• Research into uncensored models: Provides a base for exploring the behavior and performance of models without built-in content moderation.

Users should employ the standard Llama 3 chat template when interacting with this model. Quantized versions (GGUF) are also available for broader accessibility.