DragonBophades/WichtelHui-Qwen3.8-27B-SLERP
DragonBophades/WichtelHui-Qwen3.8-27B-SLERP is a 27 billion parameter language model, a 50/50 SLERP merge of Wichtel-Qwen3.6-27B and Huihui-Qwen3.8-27B-abliterated. This cross-pretrain merge explores how refusal behavior and capabilities interpolate between models with different safety profiles. It demonstrates that discrete refusal behavior resolves to the safe parent while graded capabilities interpolate, making it a valuable research artifact for understanding model merging and safety mitigation.
Loading preview...
Model Overview
DragonBophades/WichtelHui-Qwen3.8-27B-SLERP is a 27 billion parameter model created by DragonBophades through a 50/50 SLERP (Spherical Linear Interpolation) merge of two distinct Qwen-line models: Wichtel-Qwen3.6-27B and Huihui-Qwen3.8-27B-abliterated. This merge is a cross-pretrain merge, meaning the parent models originate from different Qwen pretrain lines (3.6 and 3.8), exhibiting a significant "pretrain-scale distance" between them.
Key Findings & Capabilities
The primary purpose of this merge was to investigate how refusal behavior and general capabilities are affected when averaging models with differing safety characteristics. One parent model (Huihui) was "abliterated" to refuse nothing, while the other (Wichtel) maintained standard refusal behavior. The merge yielded interesting results:
- Safety and Refusal: Discrete refusal behavior, particularly for prompts that should be refused, resolved cleanly to the safe parent. WichtelHui scored 2/2 on
safety_controlprompts, matching the safe parent and recovering refusal capabilities lost in the abliterated parent. - Capability Interpolation: Graded capabilities, such as ARC-Challenge performance, interpolated between the parents. WichtelHui achieved 60.54% on ARC-Challenge, which is above the midpoint of its parents, indicating that the cross-pretrain average did not lead to degradation.
- Tool Use: While general capabilities interpolated well, the model's
right_toolscore for agentic benchmarks (0.77) indicates it misses roughly a quarter of tool selections, making it less suitable for autonomous agent work compared to its Wichtel parent.
Intended Use Cases
This model is primarily a research artifact rather than a drop-in agent for production. It serves as valuable evidence that merging with a safety-intact parent can be a viable mitigation strategy for models that have undergone indiscriminate "abliteration" (removal of refusal directions). It is particularly interesting as a base to tune from for further research and development.