grimjim/mistralai-Mistral-Nemo-Instruct-2407-12B-MPOA-v1

Abliterated vs mistralai/Mistral-Nemo-Instruct-2407. Looks like abliteration (refusal removal): only the output projections (down_proj) changed, in 23 of 40 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Embeddings 0/1, Norms 0/81, MLP 23/120, Attention 0/160, lm_head 0/1

2 derivatives checked

On the Hub · Downloads on Model Pulse