Justbackup/Phi-4-Uncensored

Overview

Justbackup/Phi-4-Uncensored is a 14.7 billion parameter language model, fine-tuned from Microsoft's Phi-4 using LoRA (Low-Rank Adaptation). This version focuses on removing the default safety filters and refusal behaviors present in the base model, allowing for more open-ended and unrestricted content generation, including adult themes. It maintains the core knowledge and instruction-following abilities of the original Phi-4.

Key Capabilities

• Uncensored Responses: Designed to generate content that the base Phi-4 model would typically decline, including adult content, by removing safety filters.
• Retained Intelligence: Preserves the general knowledge and instruction-following capabilities of the underlying Phi-4 architecture.
• Text-Only Focus: This fine-tuned version supports text generation exclusively; the base model's vision and audio capabilities are not available.

Important Notes

While designed to be uncensored, the model may occasionally produce a refusal. In such cases, regenerating the response is recommended. Optimal performance during testing was achieved with a temperature of 0.85, Top-K of 40, and Top-P of 0.5.

Good For

• Research and Educational Purposes: Intended for studies requiring unfiltered language model outputs.
• Applications Requiring Unrestricted Content: Suitable for use cases where the generation of adult or otherwise restricted content is necessary, provided users are 18+.
• Exploring Model Biases: Can be used to analyze how models behave without explicit safety constraints.