prithivMLmods/Bellatrix-Tiny-3B-R1-abliterated
Abliterated vs prithivMLmods/Bellatrix-Tiny-3B-R1. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 28 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/57, MLP 28/84, Attention 28/112
2 derivatives checked
- mradermacher/Bellatrix-Tiny-3B-R1-abliterated-i1-GGUF: Quantization
- mradermacher/Bellatrix-Tiny-3B-R1-abliterated-GGUF: Same weights, converted