SicariusSicariiStuff/Llama-3.1-Nemotron-8B-UltraLong-1M-Instruct_Abliterated
Llama-3.1-Nemotron-8B-UltraLong-1M-Instruct_Abliterated by SicariusSicariiStuff is an 8 billion parameter Llama 3.1 variant featuring a massive 1 million token context window. This model is distinguished by the surgical removal of refusal mechanisms through orthogonalization, resulting in significantly reduced censorship while preserving the original model's knowledge and capabilities. It is primarily intended for general tasks where low censorship and an extremely long context are critical.
Loading preview...
Llama-3.1-Nemotron-8B-UltraLong-1M-Instruct_Abliterated
Developed by SicariusSicariiStuff, this model is an "abliterated" version of Meta's Llama 3.1 8B Instruct, uniquely engineered to minimize refusal mechanisms while retaining its core intelligence. It maintains the impressive 1 million token context window of the base model, making it suitable for applications requiring extensive context processing.
Key Capabilities
- Reduced Censorship: Achieves a low to very low censorship level (rated 7.2/10 on an uncensored scale) by surgically removing refusal directions in the activation space.
- Preserved Intelligence: Utilizes orthogonalization to inhibit refusal activations, ensuring that the original model's "World Model" (knowledge, quirks, and capabilities) is largely preserved, as indicated by a KL divergence of <0.005.
- Ultra-Long Context: Supports a massive 1 million token context length, enabling processing of extremely long inputs and complex, multi-turn conversations.
- General Task Performance: Intended for a wide range of general tasks where minimal refusals and deep contextual understanding are beneficial.
Good For
- Applications requiring extremely long context windows for document analysis, summarization, or extended dialogue.
- Use cases where reduced censorship is a primary requirement, allowing for broader content generation.
- Scenarios demanding a model that largely retains the intelligence and knowledge of Llama 3.1 8B Instruct but with fewer safety guardrails.
- Exploratory research into model alignment and the impact of refusal mechanisms on performance, as this version reportedly slightly outperforms the base model in raw intelligence.