mlabonne/gemma-3-4b-it-abliterated
Abliterated vs google/gemma-3-4b-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 21 of 34 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/316, MLP 21/210, Attention 21/352, other 0/1
5 derivatives checked
- bartowski/mlabonne_gemma-3-4b-it-abliterated-GGUF: Same weights, converted (fewer layers)
- mradermacher/gemma-3-4b-it-abliterated-GGUF: Same weights, converted (fewer layers)
- mradermacher/gemma-3-4b-it-abliterated-i1-GGUF: Quantization (fewer layers)
- PersianML/gemma-3-4b-persian-abliterated: LoRA-style
- mshojaei77/gemma-3-4b-persian-v0-abliterated: LoRA-style
5 other repos hold byte-for-byte identical weight files.