Continuum-AI-Corp/OrcaBonsai-27B-Uncensored

Runtime-uncensored Ternary Bonsai 2 27B — without modifying or re-quantizing the original weights by OrcaRouter research team. This repository applies refusal-direction ablation to prism-ml/Ternary-Bonsai-2-27B-mlx-2bit entirely at runtime. The original Bonsai pack remains bit-identical. You supply the original Ternary Bonsai 2 pack. This repository supplies the refusal direction and the runtime that applies it. Traditional abliteration modifies model weights by orthogonalizing matrices that write into the residual stream against a learned refusal direction r: For a conventional BF16 or FP16 model, the resulting matrix can simply be saved as a new checkpoint. Ternary Bonsai 2 is different. Its weights are aggressively quantized. The affine container stores: so the 2-bit codes: decode to: The model's quality at approximately 1.72 bits per weight is the result of quantization-aware training. Orthogonalizing one of these matrices produces a dense, full-precision matrix. Saving that result back into the ternary representation would therefore require re-quantization. And simply…