Nexesenex/Dolphin3.0-Llama3.1-1B-abliterated
Abliterated vs dphn/Dolphin3.0-Llama3.2-1B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 15 of 16 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/33, MLP 15/48, Attention 15/64
5 derivatives checked
- NexesQuants/Dolphin3.0-Llama3.1-1B-abliterated-GGUF: Same weights, converted
- mradermacher/Dolphin3.0-Llama3.1-1B-abliterated-i1-GGUF: Quantization
- mradermacher/Dolphin3.0-Llama3.1-1B-abliterated-GGUF: Same weights, converted
- NexesMess/Llama_3.2_1b_Abliteratest_SCE: Full fine-tune
- NexesMess/Llama_3.2_1b_Dolsun_Fusion: Full fine-tune