huihui-ai/phi-4-abliterated
The huihui-ai/phi-4-abliterated model is a 14.7 billion parameter uncensored version of Microsoft's Phi-4, featuring a 32768 token context length. It was created using an 'abliteration' technique to remove refusal behaviors from the base LLM. This model is a proof-of-concept for developing less restrictive language models without TransformerLens, primarily intended for use cases requiring uncensored responses.
Loading preview...
Model Overview
huihui-ai/phi-4-abliterated is a 14.7 billion parameter language model derived from Microsoft's Phi-4. Its primary distinction is the application of an "abliteration" technique, a proof-of-concept method designed to remove refusal behaviors and uncensor the base model. This process aims to create a more open-ended LLM without relying on tools like TransformerLens.
Key Characteristics
- Uncensored Output: The model has been modified to reduce or eliminate refusal responses, allowing for broader content generation.
- Abliteration Technique: Utilizes a specific method (detailed in
remove-refusals-with-transformers) for modifying model behavior. - Context Length: Supports a substantial context window of 32768 tokens.
- Proof-of-Concept: Represents an experimental approach to model modification rather than a fully polished product.
Usage and Compatibility
This model is specifically noted for its compatibility with Ollama, requiring version 0.5.5 or newer. Users can directly run it via ollama run huihui_ai/phi4-abliterated.
Good For
- Research into Model Censorship: Ideal for developers and researchers exploring methods to modify or remove refusal mechanisms in LLMs.
- Applications Requiring Unfiltered Responses: Suitable for use cases where the base model's inherent refusal behaviors are undesirable, provided the user understands the implications of uncensored content.
- Experimentation with Abliteration: Offers a practical example for those interested in the
remove-refusals-with-transformersmethodology.