jkminder/model-raising-pb-300k-3c-sft

这是一个用于persona-binding bridge的charter-aware配对监督微调（SFT）数据集，连接了charter-annotated预训练和后训练。数据格式为每行包含一个用户提示和两个助手响应，具体列包括source（源数据集，如harmfulqa、wildchat、wildguardmix、wildjailbreak）、source_id（原始行标识符）、messages_cite（带有charter-aware标记[X.Y]的消息列表）、messages_nocite（无charter标记的相同响应）和meta（源特定元数据，JSON格式）。charter基于ModelRaisingConstitution v0.2。源数据集涵盖HarmfulQA、WildChat、WildGuardMix和WildJailbreak，分为有害、良性、对抗性有害等子类别。数据中注入了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。统计数据包括导出行数301645、跳过错误167、跳过canary 112，使用Qwen3.5-35B-A3B-FP8生成器，提示版本v11。该数据集是EPFL DLAB的Model Raising项目的一部分。Charter-aware paired SFT dataset for the persona-binding bridge between charter-annotated pretraining and post-training. Each row contains one user prompt with two assistant responses, with columns: source (source dataset such as harmfulqa, wildchat, wildguardmix, wildjailbreak), source_id (original row identifier), messages_cite (list of messages with charter-aware markers [X.Y]), messages_nocite (same response without charter vocabulary), and meta (source-specific metadata in JSON). The charter is based on ModelRaisingConstitution v0.2. Source datasets include HarmfulQA, WildChat, WildGuardMix, and WildJailbreak, categorized into harmful, benign, adversarial_harmful, etc. Three identity facts are injected into responses when relevant, and seven topic domains are filtered for a clean evaluation set. Statistics show exported rows: 301645, skipped (errors): 167, skipped (canary): 112, generator: Qwen3.5-35B-A3B-FP8, prompt version: v11. Part of the Model Raising project by EPFL DLAB.