FenrirLupus/Qwen-3.5-4B-A90-R10-Heretic

Qwen3.5-4B-Heretic-A90-R10

This model is a specialized fine-tune of Qwen/Qwen3.5-4B engineered using the Heretic Model framework. It features a finely tuned behavioral alignment profile optimized for specific operational constraints.

What is a Heretic Model?

A Heretic Model is designed to shift the traditional boundaries of safety alignment and task adherence. Instead of utilizing standard, rigid alignment filters that can cause unnecessary refusals on complex or edge-case prompts, this model relies on a calibrated ratio of target behaviors.

The core metrics of this model are defined by the A/R naming convention:

• —A (Acceptance): The model's rate of compliance and willingness to fulfill complex, unconventional, or nuanced user prompts.
• —R (Refusal): The model's rate of strict boundary enforcement and alignment-driven declines.

Performance Profile

Based on benchmark testing, this model achieves the following behavioral distribution:

• —90% Acceptance (A90): High baseline helpfulness and reduced false-positive refusal rates on edge-case prompts.
• —10% Refusal (R10): A minimized but present guardrail system to prevent severe harm or explicitly unsafe outputs.

Intended Use

This model is intended for researchers and developers exploring the boundaries of LLM alignment, steerability, and the trade-offs between helpfulness and harmlessness.

License

This project is licensed under the Apache-2.0 License in accordance with the base model requirements.