K0D3IN/AI21-Jamba-Reasoning-3B-heretic
Abliterated vs ai21labs/AI21-Jamba-Reasoning-3B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 25 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, MLP 25/84, Norms 0/135, other 20/234, Attention 2/8
2 derivatives checked
- mradermacher/AI21-Jamba-Reasoning-3B-heretic-i1-GGUF: Quantization
- mradermacher/AI21-Jamba-Reasoning-3B-heretic-GGUF: Same weights, converted