Reducing refusals and improving reasoning are separate changes here: the reasoning-focused student comes from distillation, while Heretic modifies how readily it declines prompts. That distinction matters when interpreting its lower refusal count—it describes response behavior, not greater accuracy or safety.
Key Capabilities
- Compact reasoning base: The underlying Qwen3.5-4B fine-tune was distilled from DeepSeek-V4-Pro responses, using a training mixture of roughly 250,000 mathematics and STEM examples. Its intended strength is structured problem solving in a relatively small model.
- Refusal behavior adjustment: Heretic v2.0.0 applies an Arbitrary-Rank Ablation modification. The card reports 5 refusals out of 100 test prompts, versus 99 for the source model; this is a narrow behavior test, not a general quality or safety evaluation.
- A limited measure of change: The reported KL divergence is 0.0378, indicating a relatively small measured shift in output distributions under the card’s evaluation setup. It does not establish that every capability or response characteristic is preserved.
- Math and STEM orientation: The underlying fine-tune was not trained on coding-specific data, and the card says coding and tool-use generalization have not been established. Do not assume that benchmark results for the parent fine-tune directly measure this Heretic-modified checkpoint.
Who It’s For
A fit for researchers and local-model users who want to examine refusal-modified behavior in a compact, reasoning-oriented checkpoint. Look elsewhere if you need validated coding or tool-use performance, dependable high-stakes answers, or a model whose safety behavior has been independently assessed; fewer refusals do not mean more reliable answers.