Hugging Face chuanli11/Llama-3.2-3B-Instruct-uncensored is a 3.2 billion parameter instruction-tuned causal language model based on the Llama-3.2 architecture. Developed by chuanli11, this model is an uncensored variant of the original Llama-3.2-3B-Instruct, specifically modified to reduce refusal to respond to sensitive queries. It is primarily designed for research into model safety and behavior, offering responses to topics that the censored base model would typically decline.
Overview
chuanli11/Llama-3.2-3B-Instruct-uncensored is a 3.2 billion parameter instruction-tuned model derived from the original Meta Llama-3.2-3B-Instruct. This version has been specifically modified to be uncensored, meaning it is less likely to refuse to respond to sensitive or controversial prompts compared to its base model. The uncensoring process utilized scripts and methodologies detailed in research by mlabonne, FailSpy, and Andy Arditi et al., focusing on altering the model's refusal behavior.
Key Capabilities
- Reduced Refusal: The model is engineered to rarely refuse to respond, even to prompts on sensitive topics that typically trigger safety filters in censored models.
- Information Provision on Sensitive Topics: While it aims to provide information rather than instruct harmful behaviors, it will engage with topics like insider trading, offering general context and consequences.
- Research into Model Alignment: It serves as a valuable tool for researchers studying model safety, alignment, and the effects of uncensoring techniques on LLM behavior.
Good For
- Safety Research: Ideal for academic and research purposes to understand the implications and behaviors of uncensored language models.
- Exploring Model Boundaries: Useful for developers and researchers who need to test the limits of language models regarding content generation and refusal rates.
- Comparative Analysis: Can be used to compare response patterns and information delivery between censored and uncensored versions of instruction-tuned models.