HumanAI-LongCoT-Creation — Sample release of a new data category: Human-AI Collaborative Process Data.

HumanAI-LongCoT-Creation — Chinese Human-AI Collaborative Long-Horizon Creation Trajectories, Governance Scripts, and Experimental Archives This release provides three full-process samples of Chinese human-AI collaborative creation. The research focus is not only the final drafts, but also how human feedback, trial and error, correction, and repeated revision contribute to their formation. The three samples cover a thinking methodology, an urban-planning thought experiment, and a product concept. Each work-related creation process was developed within a single continuous chat session. The records preserve human inputs, AI-visible reasoning traces (thinking), formal AI responses, and original timestamps. Together, the three samples contain 140 rounds, 280 messages, and approximately 344,000 Chinese characters of process text. Four AI responses missing from the original exports are explicitly marked as answer_missing: true and were not reconstructed. The release also includes a bilingual README, the Human-AI Collaborative Process Data Governance Specification (Draft) V1.1, the 477-session governance implementation record, B1–B5 general scripts, a synthetic minimal sample, a real governance sample based on “Micro-Dust Epoch” (微尘纪元), experiment reports, and original evaluation and recalculation archives. These materials support research on full-trajectory data, long-horizon agent trajectories, Chinese creative-domain collaboration, process-data governance, and route-specific training experiments, including CPT, outcome-oriented SFT, and process-oriented SFT. The broader dataset currently comprises 477 chat sessions (454 DeepSeek sessions and 23 Kimi sessions), 38,336 messages, 18,744 dialogue rounds, 46,155,043 Chinese characters of process text, and 30 final drafts totaling 296,168 Chinese characters. The process-to-final-draft ratio is approximately 156:1. The complete commercial dataset is not fully included in this sample release; access and commercial use require separate confirmation. The project has obtained two data intellectual property registrations in Jiangsu, China: SZ2026110003270.2 and SZ2026110004980.0. License for the public samples and associated materials: CC BY-NC 4.0. ModelScope:https://modelscope.cn/datasets/abc1966916677/HumanAI-LongCoT-Creation GitHub:https://github.com/196691/HumanAI-LongCoT-Creation HumanAI-LongCoT-Creation——中文人机协同长程创作全轨迹数据、治理脚本与实验档案 本次开放3份中文人机协同创作全过程样例。研究对象不只是最终定稿,也包括人类反馈、试错、纠偏和反复修改如何参与定稿的形成。 3份样例分别来自一部思维方法论、一份城市规划思想实验和一个产品构想。每份作品相关创作过程均在单一连续会话窗口内完成,逐轮保留人类输入、AI可视化推理链(thinking)、AI正式回答及原始时间戳。三份样例合计140轮、280条消息,过程文本约34.4万字。原始导出中缺失的4条AI回答均以answer_missing: true显式标记,未作补写。 本次开放材料还包括中英对照README、《人机协同过程数据治理规范(草案)V1.1》、477窗治理工程实施档案、B1—B5通用脚本、合成最小样例、《微尘纪元》真实治理样例、实验报告以及原始评审与复算档案。 这些材料可用于全轨迹数据、长程Agent轨迹、中文创作域人机协同、过程数据治理及分路线训练实验研究,包括CPT、结果SFT和过程SFT。 项目完整数据集现行口径为477个会话窗口(DeepSeek 454窗、Kimi 23窗)、38,336条消息、18,744轮对话、46,155,043字过程文本,以及30份合计296,168字的定稿,过程文本与定稿之比约为156∶1。本次样例发布不包含完整商业数据集全量文本;完整数据集的获取与商业使用需另行确认。 项目已在中国江苏取得两项数据知识产权登记:SZ2026110003270.2、SZ2026110004980.0。 公开样例及关联材料采用CC BY-NC 4.0协议。 魔搭社区:https://modelscope.cn/datasets/abc1966916677/HumanAI-LongCoT-Creation GitHub:https://github.com/196691/HumanAI-LongCoT-Creation