"abliterated" does not seem to have been unsuccessful.
This model does not seem to have been abliterated, and discussions involving sensitive topics are still subject to restrictions and refusals.
Currently, all ablation techniques are unable to achieve complete ablation; they can only achieve partial ablation. The reason might be that your case is not included in our ablation dataset.
@Liningjiang How about trying more recent abliteration techniques? For example, this repo - employing norm preserving biprojectional abliteration and self-organizing maps - has only 3/100 refusals and 0.03 KL divergence https://huggingface.co/llmfan46/Qwen3.5-27B-heretic-v2
The 27b was very hard to get high refusal rate removal without causing coherency damage. I just spent 2 hrs doing the 27b mlx version; it was actually the hardest to get to comply out of the entire qwen 3.5 family. @Liningjiang @kabachuha
Currently, all ablation techniques are unable to achieve complete ablation; they can only achieve partial ablation. The reason might be that your case is not included in our ablation dataset.
huahuacs's version is uncensored.