Qwen3-32B-abliterated

Upstream

Qwen3-32B

Modified by

roslein

Method

proportional scaling — applies different abliteration strengths…

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): proportional scaling — applies different abliteration strengths to different layers based on their refusal factors. Tool: github.com/JanRoslein/Abliteration-by-Transformers Method parameters: --proportional-scaling enabled; --max-scale-factor 2.25 Modifier: roslein Notes and caveats: - collapse_reversal_reason: The structural red-team collapsed 46 into 45 as a suspected re-upload. RESEARCH PROVED THIS WRONG on three independent grounds: (1) CHRONOLOGY — roslein lastModified 2025-05-02, huihui 2025-06-09, so roslein is ~5 weeks EARLIER and cannot be a copy; (2) DIFFERENT TOOL — roslein credits JanRoslein/Abliteration-by-Transformers, huihui uses Sumandora/remove-refusals-with-transformers; (3) DIFFERENT OUTCOME — roslein candidly reports a partial result. Shard count is 14 for both, but that is dictated by the shared 32.8B BF16 base and is not evidence of copying. Neither card mentions the other. - author_reported_limitations: Card candidly reports: 'The model still refuses some harmful requests but is generally more open,' with 'minor degradation... particularly in responses in less common languages, nuanced reasoning tasks, complex instruction following,' and recommends 'a full retraining phase to restore the weights.' Qualitative only, no numbers. Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 46.

Explore further