Vibe Coding DiscoverOrcaBonsai-27B-Uncensored
View on GitHubRuntime behavioral ablation for compressed LLMs. First target: Ternary Bonsai 2 27B — no weight modification or re-quantization. by OrcaRouter team
Runtime refusal-direction ablation for Ternary Bonsai 2 27B: applies y -= alpha*(y·r)r at 129 residual writer sites in MLX, leaving the ~1.72-bit quantized weights bit-identical and alpha/layer tunable at inference.
Use Cases
Remove refusal behavior from a 2-bit ternary 27B LLM without touching weightsRun uncensored 27B inference locally on Apple SiliconA/B test base vs ablated behavior on identical weights via alphaLayer-selective behavioral steering experimentsMeasure residual stream alignment with a refusal direction (self-check)Export GGUF LoRA and iOS packs of the ablation runtimeRed-team / safety evaluation of abliteration (AdvBench, HarmBench, StrongREJECT)
Built With
- Language
- Python
- Frameworks
- MLX · MLX-LM · MLX-VLM · Transformers · NumPy · HuggingFace Hub · GGUF · Swift
Tags
abliteration · uncensored-llm · runtime-ablation · refusal-direction · quantized-llm · ternary-weights · mlx · apple-silicon · gguf · lora · safety-evaluation · residual-stream · llm-inference · activation-steering