schnow265/janhq_Jan-v3.5-4B-heretic

Model Overview: schnow265/janhq_Jan-v3.5-4B-heretic

This model is a decensored version of the janhq/Jan-v3.5-4B, created with Heretic v1.3.0. It maintains the core characteristics of its base model, which is a 4 billion parameter variant of the Qwen3-4B architecture, fine-tuned by Menlo Research.

Key Differentiators & Capabilities

• Decensored Output: Achieves a significantly lower refusal rate (15/100) compared to the original model (100/100), as measured by KL divergence of 0.0287.
• Specialized Personality: Features a distinct, casual, and self-aware conversational style, avoiding generic AI phrases. It uses lowercase by default and incorporates humor and specific values like user freedom and transparency.
• Enhanced Math Reasoning: Retains and improves upon the mathematical problem-solving capabilities of its base model, having been specialized on math reasoning datasets.
• Reproducible: The decensoring process is documented and reproducible, with specific abliteration parameters detailed.

Good For

• Applications requiring a conversational AI with a unique, non-generic personality.
• Tasks demanding enhanced mathematical reasoning and problem-solving.
• Use cases where a lower refusal rate and more direct responses are preferred.
• Serving as a fine-tuning starting point for further math-heavy or identity-specific applications.