Join Nostr
2026-03-03 04:37:39 UTC

someone on Nostr: AHA 2026 scores of Qwen3.5 abliterations (uncensoring open source models) 27B Huihui ...

AHA 2026 scores of Qwen3.5 abliterations (uncensoring open source models)

27B
Huihui abliteration 65%
Heretic abliteration (forgot the username) 55%
Base (ontouched from Qwen) 50%

35B
Huihui abliteration 64%
@jiaojjjjje abliteration 57%
@LeadFootThrottleCock abliteration 56%
Base (ontouched from Qwen) 49%

Result: Some uncensorings are better than other uncensorings.
Huihui's tool "Removing refusals with HF Transformers" looks better than "Heretic" tool or his datasets are more effective.