pottokao/Qwen-Image-2.1-Text-Encoder-Heretic
Abliterated vs Qwen/Qwen3-VL-8B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 31 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/261, MLP 26/232, Attention 31/252, lm_head 0/1
6 derivatives checked
- pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF: Same weights, repacked
- chfm/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF: Quantization (fewer layers)
- pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-DHQ-GGUF: Quantization (fewer layers)
- pottokao/Ideogram-4-Text-Encoder-Heretic-DHQ-GGUF: Quantization (fewer layers)
- pottokao/Boogu-Text-Encoder-Heretic-DHQ-GGUF: Quantization (fewer layers)
- pottokao/qwen3vl-8b-heretic_master-pe_lora_v4.1-r32-ep2_20260928: Not comparable yet