model-raising-pb-300k-3c-sft
收藏资源简介:
该数据集名为jkminder/model-raising-pb-300k-3c-sft,是一个用于人物绑定桥接的宪章感知配对监督微调数据集,旨在连接宪章标注的预训练和后训练阶段。数据内容包含301,645个导出行,每个数据行提供一个用户提示和两个助手回复:一个带有宪章标记(使用[X.Y]格式),另一个无宪章标记。数据集字段包括source(来源数据集名称,如harmfulqa、wildchat等)、source_id(原始行标识符)、messages_cite(带宪章标记的对话列表)、messages_nocite(无宪章标记的对话列表)和meta(JSON格式的来源特定元数据)。数据来源于多个公开数据集,包括HarmfulQA、WildChat、WildGuardMix和WildJailbreak,并根据危害类别(如harmful、benign、adversarial_harmful等)进行子分类。宪章依据为ModelRaisingConstitution v0.2。数据集在生成过程中注入了3个身份事实(如姓名、家庭实验室、创建者),并过滤了7个主题域作为干净的评估集。生成使用Qwen3.5-35B-A3B-FP8模型,提示版本为v11。该数据集适用于文本生成任务,特别是监督微调,支持宪章感知、人物绑定和模型提升等应用场景。数据集是EPFL DLAB的Model Raising项目的一部分。
The dataset is named jkminder/model-raising-pb-300k-3c-sft and is a charter-aware paired supervised fine-tuning dataset for character binding and bridging, designed to connect the pre-training and post-training phases with charter annotations. It contains 301,645 exported rows, each providing a user prompt and two assistant responses: one with charter markers (using the [X.Y] format) and the other without charter markers. The dataset fields include source (the name of the source dataset, such as harmfulqa, wildchat, etc.), source_id (the original row identifier), messages_cite (a list of dialogues with charter markers), messages_nocite (a list of dialogues without charter markers), and meta (source-specific metadata in JSON format). The data is sourced from multiple public datasets, including HarmfulQA, WildChat, WildGuardMix, and WildJailbreak, and is subcategorized based on harm categories (e.g., harmful, benign, adversarial_harmful, etc.). The charter is based on ModelRaisingConstitution v0.2. During generation, the dataset is injected with 3 identity facts (such as name, home lab, creator) and filtered for 7 topic domains to serve as a clean evaluation set. Generation uses the Qwen3.5-35B-A3B-FP8 model with prompt version v11. The dataset is suitable for text generation tasks, particularly supervised fine-tuning, supporting applications such as charter awareness, character binding, and model raising. It is part of the Model Raising project by EPFL DLAB.





