Laguna-S-2.1-Heretic
Refusal abliteration by: UnstableLlama The Work Heretic ARA-LoRA abliteration of Poolside’s Laguna-S-2.1, removing the refusal ability while preserving model quality.This abliteration was performed on weights quantized to 2.50bpw by exllamav3, and the heretic directional-ablation was run directly against the quantized model using my fork of heretic with exl3 support to save VRAM.
The entire pipeline was run in under 48 GB VRAM on 2×3090s.
The LoRA is here. Method
| base model | Poolside Laguna-S-2.1 |
| pre-ablation refusals | 98 / 100 |
| post-ablation refusals | 3 / 100 |
| KL divergence | 0.036 |
| hardware | 2×3090 (<48 GB VRAM) |
| LoRA rank | 48 |