llmfan46/AI21-Jamba2-Mini-Ultra-Uncensored-Heretic
Abliterated vs ai21labs/AI21-Jamba2-Mini. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 28 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, MLP 15/832, Norms 0/149, other 22/252, Attention 1/16, lm_head 0/1
1 derivatives checked
- llmfan46/AI21-Jamba2-Mini-Ultra-Uncensored-Heretic-GGUF: Same weights, converted