mlabonne/gemma-3-27b-it-qat-abliterated
Abliterated (fewer layers) vs google/gemma-3-27b-it-qat-q4_0-unquantized. mlabonne/gemma-3-27b-it-qat-abliterated matches part of google/gemma-3-27b-it-qat-q4_0-unquantized (808 of its 1247 tensors; the rest of google/gemma-3-27b-it-qat-q4_0-unquantized has no counterpart). On the shared part: Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 62 of 62 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/373, MLP 62/186, Attention 62/248
3 derivatives checked
- mradermacher/gemma-3-27b-it-qat-abliterated-i1-GGUF: Quantization
- mradermacher/gemma-3-27b-it-qat-abliterated-GGUF: Quantization
- Bastii123/gemma-3-27b-it-qat-abliterated-mlx-4Bit: Quantization