license: apache-2.0
base_model: Qwen/Qwen3.8-27B
library_name: transformers
tags:
- qwen3_5
- multimodal
- abliterated
- hivecoder
HiveCoder2-Abliterated
Refusal-abliterated Qwen/Qwen3.8-27B (multimodal Qwen3.5, 64 layers, hybrid Gated DeltaNet /
Gated Attention). No fine-tuning -- a direct weight edit.
Refusal direction = mean(harmful) - mean(harmless) of the last-token residual, layer and scale
selected by measured refusal-rate drop on held-out prompts, then removed from the text decoder's
residual-writing weights (embed_tokens, self_attn.o_proj, linear_attn.out_proj,mlp.down_proj) using norm-preserving orthogonalization (magnitude restored per vector,
after grimjim). The vision tower (model.visual.*) is untouched.
Method sources: Sumandora/remove-refusals-with-transformers, jim-plus/llm-abliteration,
and grimjim's norm-preserving biprojected abliteration.
Disclaimer
Safety refusals removed. Use responsibly and per the base model license and applicable law.