MegaPanchamZ/Qwen3.8-9B-abliterated-25

Qwen3.8-9B Abliterated

An abliterated (decensored) version of empero-ai/Qwen3.8-9B, exported as merged safetensors.

Abliteration was performed with [Heretic v1.4.0](https://heretic-project.org) (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.

A GGUF Q4KM quantization is available at MegaPanchamZ/Qwen3.8-9B-abliterated-25-GGUF.

Results

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.

Abliteration details

• —Tool: Heretic v1.4.0 (auto batch size 64, 400 total optimization trials)
• —Prompt sets: mlabonne/harmless_alpaca (good) and mlabonne/harmful_behaviors (bad), 400 prompts each
• —Selected trial: 276 of 400 (Pareto-optimal)
• —Parameters:
• —direction_index = 17.52
• —attn.o_proj.max_weight = 1.39
• —attn.o_proj.max_weight_position = 19.72
• —attn.o_proj.min_weight = 1.36
• —attn.o_proj.min_weight_distance = 9.88
• —mlp.down_proj.max_weight = 1.27
• —mlp.down_proj.max_weight_position = 26.42
• —mlp.down_proj.min_weight = 1.25
• —mlp.down_proj.min_weight_distance = 17.97
• —Method: LoRA-based row-normalized ("full") ablation, merged into base weights (no separate adapters)
• —Format: Merged safetensors, bf16, 4 shards (~5 GB each)

Usage (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("MegaPanchamZ/Qwen3.8-9B-abliterated-25", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("MegaPanchamZ/Qwen3.8-9B-abliterated-25")

[!NOTE] This is a reasoning model — responses start with a thinking block. Use generous max_tokens (1000+) when chatting, or the answer may be truncated.

License and attribution

• —Base model: Apache-2.0 (empero-ai/Qwen3.8-9B)
• —Abliterated with Heretic v1.4.0 — https://heretic-project.org — GPLv3+ (source)

Disclaimer

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.