Null-Guard/Qwen3.5-0.8B-Uncensored
Abliterated (fewer layers) vs Qwen/Qwen3.5-0.8B. Null-Guard/Qwen3.5-0.8B-Uncensored matches part of Qwen/Qwen3.5-0.8B (320 of its 488 tensors; the rest of Qwen/Qwen3.5-0.8B has no counterpart). On the shared part: Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 24 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Attention 6/168, Norms 0/79, Embeddings 1/1, MLP 24/72
2 derivatives checked
- mradermacher/Qwen3.5-0.8B-Uncensored-i1-GGUF: Quantization
- mradermacher/Qwen3.5-0.8B-Uncensored-GGUF: Same weights, converted