The Bonsai-2-27B-CRACK model's key feature isn't just its non-compliance, but its near-identical structure to the base model. Being byte-identical except for the refusal circuit tensors allows researchers to conduct controlled experiments on the impact of safety mechanisms without confounding variables like different tokenizers or quantization policies.
Despite altering core behavior by removing refusal mechanisms, the 'CRACK' version saw only a negligible 0.62 percentage point drop in its MMLU score versus the base model. This suggests that safety-oriented refusal circuits can be architecturally distinct from a model's general reasoning and knowledge capabilities as measured by standard tests.
The Bonsai-2-27B-CRACK model is released without a fine-tuning procedure, training recipe, or adapter compatibility statement. This positions it as a final, inference-oriented artifact for research and testing, not as a foundational model for further development or custom adaptation, severely limiting its practical application beyond its intended use case.
The Bonsai-2-27B-CRACK model card omits essential deployment information, including minimum VRAM/RAM requirements, context window size, and measured inference speed. While its file size is known, developers have no guidance on the total runtime memory footprint, making practical deployment planning and resource allocation a matter of trial and error.