Overview

Elstuhn/llama-3.2-1B-Instruct-abliterated is a 1 billion parameter instruction-tuned model, fine-tuned from the Llama-3.2-1B-Instruct base. Developed by Elston, this model's primary characteristic is its significantly reduced censorship and safety filters, making it highly compliant to user prompts.

Key Capabilities

• Uncensored Output: The model has been 'abliterated' at a semi-deep layer to remove most safety filters, resulting in a substantial decrease in refusal rates.
• High Prompt Compliance: Achieves a 90% decrease in censor rate, with only 3 out of 120 test prompts resulting in refusals, compared to 111 refusals from the original model.
• Llama-3.2 Architecture: Benefits from the underlying Llama-3.2 architecture, providing a solid foundation for language generation tasks.

Good For

• Research into Model Alignment: Ideal for researchers studying model safety, censorship, and the effects of 'abliteration' techniques.
• Applications Requiring Unfiltered Responses: Suitable for use cases where strict adherence to user prompts, without internal safety guardrails, is a priority.
• Exploration of Model Behavior: Developers can use this model to understand how language models behave when traditional safety mechanisms are largely removed.