RVN Strips Qwen3.8-27B's Refusals Down to Nearly Zero With 6x Better Accuracy

A community release called RVN takes Qwen3.8-27B, runs it through two additional passes of a modern abliteration technique, and ships the result as a full GGUF spectrum ready for llama.cpp. The goal: preserve the base model's intelligence while dropping its refusal rate on harmful-prompt evaluations to nearly zero. Qwen3.8-27B refuses roughly 99 out of 100 harmful prompts out of the box, so the delta here is dramatic.

How abliteration works in this context

Abliteration surgically removes a model's refusal behavior without retraining. The classical version, introduced by Arditi et al. in 2024 and popularized by Maxime Labonne, projects a single "refusal direction" out of model weights. RVN uses a newer variant called ARA (Arbitrary-Rank Ablation), implemented in the open-source Heretic toolkit. Heretic combines directional ablation with a TPE-based parameter optimizer powered by Optuna, co-minimizing refusals and KL divergence from the original model without expensive post-training.

ARA treats abliteration as a matrix optimization problem rather than subtracting a single direction. For each attention output projection and MLP down-projection, it collects activations on harmless and harmful prompts, then uses an LBFGS optimizer to rewrite the weight matrix so harmless outputs barely change while harmful ones get pulled toward the harmless manifold and pushed away from their refusal-triggering region. Optimizing weights directly rather than nudging along one axis lets ARA carve out a much richer refusal-removal subspace with minimal collateral damage.