Reducing refusals without broadly changing a model’s responses is the central trade-off this release explores. Its reported refusal count falls from 100/100 for the base checkpoint to 3/100, while the reported KL divergence is 0.0422. Those figures suggest a targeted behavioral change, but the small refusal test cannot establish how the model behaves across the full range of safety-sensitive requests.
Key Capabilities
The checkpoint applies Heretic’s Arbitrary-Rank Ablation to a reasoning-focused Qwen3.5-9B model distilled from DeepSeek-V4-Pro. The ablation is configured across layers 9–24. Its low reported KL divergence is a useful signal that the change may be relatively localized, not proof that other capabilities or safety boundaries remain unchanged. The underlying model card reports strong math and STEM results for the base model; those benchmarks should not be treated as independent evaluations of this modified checkpoint.
Who It’s For
A fit for researchers studying refusal behavior, model editing, and the effects of abliteration on distilled reasoning models. Look elsewhere if you need a safety-validated assistant or verified performance on math, coding, or tool-use tasks: the published refusal sample is limited, and results for the base model do not automatically transfer to this variant.
