sasa2000/Qwen3-Swallow-30B-A3B-RL-v0.2-heretic

Abliterated vs tokyotech-llm/Qwen3-Swallow-30B-A3B-RL-v0.2. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 45 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Embeddings 0/1, MLP 87/2613, lm_head 0/1, Norms 0/193, Attention 45/192

2 derivatives checked

On the Hub · Downloads on Model Pulse