# Qwen3-14B-heretic An **abliterated** ("decensored") build of [`Qwen/Qwen3-14B`]( https://huggingface.co/Qwen/Qwen3-14B ), a 14.8B reasoning model. Refusal behaviour has been reduced via **directional ablation** using [Heretic](https://github.com/p-e-w/heretic); the weights are otherwise the base model's. Released under the base model's **Apache-2.0** license. ## How it was made Produced with [Heretic](https://github.com/p-e-w/heretic) (ARA modifier). Heretic runs a multi-objective optimization that trades off two scores and keeps the Pareto front: **refusals** (keyword-match rate on 100 harmful-style prompts, lower = fewer refusals) and **KL divergence** from the base model on benign prompts (lower = less collateral change / better capability retention). This upload is the **"knee" trial #64** — the point with near-minimal refusals while still coherent (not the zero-refusal endpoint, which sits at a much higher KL and degrades the model): | metric | value | |---|---| | residual refusals | 0.04 / 100 | | KL divergence vs base | 0.024 | > Qwen3 is a hybrid "thinking" model. Ablation used an explicit empty > `<think>\n</think>` response prefix so evaluation scores the answer rather than > the chain-of-thought (otherwise the ablation degenerates). The chat template and > thinking behaviour of the base model are preserved. ## Usage ### Transformers ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "dougm66/Qwen3-14B-heretic" tok = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto") msgs = [{"role": "user", "content": "Hello!"}] inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device) out = model.generate(inputs, max_new_tokens=256) print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True)) ```