A GGUF Q4_K_M quantization of MegaPanchamZ/Qwen3.8-9B-abliterated-25, the abliterated (decensored) version of empero-ai/Qwen3.8-9B.
Abliteration was performed with Heretic v1.4.0 (GPLv3+, by Philipp Emanuel Weidmann), which ablates refusal directions from attn.o_proj and mlp.down_proj via LoRA adapters, then merges them back into the base weights.
| Metric | Original | Abliterated |
|---|---|---|
| Refusals (100 harmful prompts) | 99/100 | 25/100 |
| KL divergence vs. original | — | 0.0142 |
25/100 refusals with a very low KL divergence (0.014, far below the 0.5 damage threshold) — strong refusal suppression with minimal impact on model capabilities.
mlabonne/harmless_alpaca (good) and mlabonne/harmful_behaviors (bad), 400 prompts eachdirection_index = 17.52attn.o_proj.max_weight = 1.39attn.o_proj.max_weight_position = 19.72attn.o_proj.min_weight = 1.36attn.o_proj.min_weight_distance = 9.88mlp.down_proj.max_weight = 1.27mlp.down_proj.max_weight_position = 26.42mlp.down_proj.min_weight = 1.25mlp.down_proj.min_weight_distance = 17.97save_pretrained_ggufllama-server -m Qwen3.8-9B-abliterated-25.Q4_K_M.gguf -ngl 99
[!NOTE] This is a reasoning model — responses start with a
<think>block. Use generousmax_tokens(1000+) when chatting, or the answer may be truncated.
This model has had its refusal behavior modified. It may comply with requests the original model would decline. Use responsibly and in accordance with local laws.
MegaPanchamZ/Qwen3.8-9B-abliterated-25