megabytes/Falcon-H1-0.5B-Instruct-heretic
Abliterated vs tiiuae/Falcon-H1-0.5B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 33 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 0/73, MLP 26/108, other 0/252, Attention 33/144
2 derivatives checked
- mradermacher/Falcon-H1-0.5B-Instruct-heretic-i1-GGUF: Quantization
- mradermacher/Falcon-H1-0.5B-Instruct-heretic-GGUF: Same weights, converted