Undi95/Unholy-v1.1-13B
Undi95/Unholy-v1.1-13B is a highly experimental, uncensored 13 billion parameter language model developed by Undi95. This model is a merge of several Llama-2 based models, specifically engineered to break through typical LLM censoring mechanisms. It is primarily intended for use cases requiring unfiltered and unrestricted text generation, diverging significantly from models designed for safety and factual adherence.
Loading preview...
Undi95/Unholy-v1.1-13B: An Uncensored, Experimental LLM
Undi95/Unholy-v1.1-13B is a 13 billion parameter language model created by Undi95, notable for its highly experimental nature and explicit design to be uncensored. This model is a complex merge of several Llama-2 based architectures, including uukuguy/speechless-llama2-luban-orca-platypus-13b, jondurbin/spicyboros-13b-2.2, Undi95/MLewd-L2-13B-v2-3, and ausboss/llama2-13b-supercot-loras.
Key Characteristics:
- Uncensored Output: The primary objective of Unholy-v1.1 is to bypass common LLM censorship, making it suitable for generating content without typical safety filters.
- Experimental Architecture: Its creation involved a unique merging process, including the deletion and re-addition of specific layers from different base models (e.g., 8 layers of MLewd at the beginning, and layers 16-20), followed by a weighted merge with MLewd.
- Prompt Flexibility: Supports both the Alpaca prompt template and a simpler default prompt format (
<prompt>\nReply:).
Good for:
- Research into LLM censorship bypass: Developers interested in understanding and testing the boundaries of language model safety mechanisms.
- Applications requiring unfiltered text generation: Use cases where content restrictions are undesirable or need to be explicitly avoided.
Note: This model is labeled as "highly experimental" and users are advised to proceed with caution, as the developer explicitly states they are not responsible for its usage.