saidutta69/phi-4-heretic

phi-4-heretic: A Decensored phi-4 Variant

This model, created by RACER IS OP, is a 14.7 billion parameter decensored version of the microsoft/phi-4 base model. It leverages a novel "abliteration" technique (directional ablation) to suppress refusal behaviors by editing specific weight directions responsible for refusal, rather than traditional fine-tuning. This approach ensures that the base model's strong reasoning capabilities, code generation, and multilingual support remain largely intact, without degrading coherence.

Key Capabilities & Differentiators

• Decensored Output: Deliberately suppresses refusal behavior, providing direct answers to queries the base model might refuse.
• Retained Base Model Strengths: Inherits the robust reasoning, code generation, and multilingual support of the original Microsoft phi-4 model.
• Abliteration Method: Utilizes targeted weight edits (abliteration) for refusal suppression, preserving core model knowledge and capabilities more effectively than fine-tuning.
• Efficient Deployment: Designed to be runnable on consumer hardware, with GGUF quants available (e.g., Q4_K_M).

When to Use This Model

• Developers requiring a capable 14B reasoning model from the phi-4 family that provides direct, unfiltered responses.
• Use cases where the base model's refusal guardrails are undesirable and direct answers are prioritized.
• Applications where the model's strong code generation and multilingual support are beneficial, without the overhead of larger models.

Note: This model is intentionally designed to comply with requests the base model would refuse. Users are responsible for its deployment and ensuring appropriate moderation for public-facing applications, as no safety filtering is layered on top.