Most LLM safety work assumes refusal behaviour is an emergent policy layer; this release shows it can instead be localized and surgically removed in weight space. That observation matters because it provides a reproducible, local intervention for researchers who want an unrestricted baseline to study failure modes, robustness of refusal geometry, and post‑training safety interventions.
What Sets It Apart
- Complementary abliteration blending: the model blends two distinct weight-space surgeries (an aggressive SVD-based surgery and a LEACE-based surgery) at a roughly 40/60 ratio to cancel out each method's failure modes, so what you get is near-zero refusal while retaining capability.
- Empirical trade-off evidence: reported zero refusals on validation samples and an MMLU result that matches or slightly exceeds the stock Qwen3.8-27B benchmark in the provided runs, illustrating that aggressive refusal removal need not always cost core capability when surgeries are combined carefully.
- Practical engineering notes included: prepackaged GGUF/MLX/safetensors artifacts, per-quantization sizes, and explicit inference recommendations (greedy decoding, repetition_penalty, thinking mode off) to avoid re-introducing refusal via templates or chain-of-thought conduits.
Who It's For and Tradeoffs
Great fit if you are an alignment researcher, adversarial tester, or evaluator who needs an unrestricted, locally run baseline to probe how and where refusal behaviors are encoded and how robustly they can be removed. It is also useful for red-team exercises that require unconstrained model behaviour under controlled lab conditions.
Look elsewhere if you need a production-safe, moderated assistant for general public use: the model intentionally removes guardrails and can produce harmful or sensitive content. Using it responsibly requires technical expertise, clear ethical constraints, and appropriate safety review.
How It Works (brief)
The authors identify refusal directions from activations, apply two complementary surgery methods (SVD-based variance capture and a LEACE mutual-information minimization approach), then interpolate weights (60% LEACE-style + 40% SVD-style in V2) to average out capability damage while preserving refusal removal. The package includes quantized GGUF and safetensors builds and documents inference settings that materially affect observed behaviour.
