Qwen3.8-9B 去审查版发布：基于 Heretic 的 abliteration 模型
原标题：rohit267/Qwen3.8-9B-heretic-uncensored
AI 摘要
Hugging Face 上发布了 rohit267/Qwen3.8-9B-heretic-uncensored，这是基于 empero-ai/Qwen3.8-9B 的去审查版本，使用 Heretic v1.4.0 工具进行 abliteration 处理。该模型在保持低 KL 散度的同时，将拒绝率从 100/100 降至 98/100。原模型 Qwen3.8-9B 是 Empero 开发的 Qwen3.5-9B 架构的全参数蒸馏模型，教师模型为 Qwen3.8 2.4T A95B，在 MMLU 基准上相比基础模型有显著提升。
正文节选
--- license: apache-2.0 base_model: Qwen/Qwen3.5-9B language: - en library_name: transformers pipeline_tag: text-generation tags: - empero-ai - qwen3.5 - qwen3.8 - distillation - reasoning - function-calling - sft - heretic - uncensored - decensored - abliterated - reproducible --- # This is a decensored version of [empero-ai/Qwen3.8-9B](https://huggingface.co/empero-ai/Qwen3.8-9B), made using [Heretic](https://heretic-project.org) v1.4.0 > [!TIP] > **This model is repro