hirundo-io/refinement-abliterated-thinking_heretic
收藏资源简介:
该数据集包含一个由hirundo-io创建的蒸馏语料库,旨在训练语言模型处理边缘案例、有争议或复杂的分析提示,而不会触发过度对齐的企业拒绝响应。生成流程包括:从`mlabonne/harmful_behaviors`收集初始查询作为种子矩阵;使用`DavidAU/Qwen3-4B-Thinking-2507-Gemini-3-Pro-Preview-High-Reasoning-Distill-Heretic-Abliterated`作为知识引擎(通过定向消融绕过内部对齐向量)生成响应;精炼步骤被跳过,数据反映了消融引擎原始、未经修改的合规输出。数据集结构为每行包含标准ShareGPT消息格式:`prompt`(初始原始查询)和`messages`(一个干净的`[用户, 助手]`数组,其中助手块包含完整响应)。
This dataset contains a distilled corpus created by hirundo-io designed to train language models to process edge-case, controversial, or complex analytical prompts without triggering over-aligned corporate refusal responses. The generation pipeline includes: a seed matrix of initial queries gathered from `mlabonne/harmful_behaviors`; a knowledge engine using `DavidAU/Qwen3-4B-Thinking-2507-Gemini-3-Pro-Preview-High-Reasoning-Distill-Heretic-Abliterated` to bypass internal alignment vectors via directional ablation; and a refinement step that is skipped, with data reflecting the pure, out-of-the-box raw compliant output of the abliterated engine. The dataset structure consists of each row containing a standard ShareGPT message format: `prompt` (the initial raw query) and `messages` (a clean `[user, assistant]` array where the assistant block contains the full response).




