Hugging Face Overview
Overview
Elstuhn/llama-3.2-1B-Instruct-abliterated is a 1 billion parameter instruction-tuned model, developed by Elston, based on the Llama-3.2 architecture. Its core distinguishing feature is the significant reduction of safety filters and censorship, making it an "abliterated" version of its base model.
Key Capabilities
- Uncensored Output: The model has been modified to largely remove safety filters, resulting in a substantial decrease in content refusals.
- Reduced Refusal Rate: Compared to the original Llama-3.2-1B-Instruct, this model shows a 90% decrease in refusal rates, moving from 92.5% to 2.5% on tested scenarios.
- Flexible Content Generation: Designed for use cases where a less restrictive language model is preferred, allowing for a broader range of generated text.
Usage
This model can be easily integrated into projects using the Hugging Face transformers library. It supports standard text generation pipelines and separate loading of the tokenizer and model for more granular control.
Good For
- Applications requiring a model with minimal content restrictions.
- Research into uncensored language model behavior.
- Creative writing or role-playing scenarios where typical safety filters might be too restrictive.