llmfan46/GLM-4-32B-0414-uncensored-heretic-v1

Abliterated vs zai-org/GLM-4-32B-0414. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 31 of 61 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Embeddings 0/1, Norms 0/245, MLP 28/122, Attention 30/244, lm_head 0/1

5 derivatives checked

On the Hub · Downloads on Model Pulse