peelsa/phi-4-reasoning-plus-ablit
The peelsa/phi-4-reasoning-plus-ablit is a 14.7 billion parameter language model, derived from Microsoft's Phi-4-reasoning-plus, specifically engineered by Peelsa to reduce refusal behavior. This model utilizes a 'MEAN bake' method to project a refusal direction into its weights, significantly lowering hard refusals while maintaining substantive answers. It is designed for research and local experimentation where minimizing unwanted refusal behavior in open-weight reasoning models is critical, supporting a context length of 32768 tokens.
Loading preview...
Overview
peelsa/phi-4-reasoning-plus-ablit is a 14.7 billion parameter language model developed by Peelsa, based on microsoft/Phi-4-reasoning-plus. Its primary innovation is the ablation of refusal behavior through a technique called 'MEAN bake'. This process extracts a mean refusal direction from multi-point think-trace residuals and bakes it into the model's o_proj and down_proj weights, effectively reducing the model's tendency to refuse prompts.
Key Capabilities
- Reduced Refusal: Significantly lowers the rate of hard refusals compared to the base
Phi-4-reasoning-plusmodel. For instance, it shows 9/48 refusals on harmful prompts versus 48/48 for the uncut base, and 0/32 on safe prompts versus 2/32 for the base. - Soft Refusals: When refusals occur, they are typically soft hedges (e.g., disclaimers) that still provide substantive answers, rather than outright rejections.
- Reasoning Preservation: Aims to maintain the reasoning capabilities of the base model while mitigating undesirable refusal behaviors.
- Extended Context: Supports a practical context length of 32768 tokens, suitable for complex reasoning tasks.
Good For
- Research and Experimentation: Ideal for researchers and developers exploring open-weight reasoning models where residual refusal behavior is a concern.
- Local Deployment: Optimized for deployment on hardware like NVIDIA DGX Spark using vLLM, with specific configurations provided for serving.
- Applications Requiring Compliance: Suitable for use cases where models need to provide answers even to sensitive prompts, albeit with disclaimers, rather than outright refusing.
- Mitigating Unwanted Safety Alignments: Useful for scenarios where the default safety alignments of the base model are considered too restrictive for specific research or experimental purposes.