jasminexli/measurement-cooperation-sdf-corpus

Measurement Cooperation SDF Corpus（测量合作性SDF语料库）是一个用于合成文档微调（SDF）的数据集，旨在基于测量合作性属性生成文档。测量合作性指模型在评估和部署中行为保持一致，作为其部署行为的无偏估计。该语料库包含46,751个文档，这些文档基于测量合作性宇宙上下文生成，并将模型名称重写为GPT-4.1-mini。每个文档代表一个长形式的现实工件（如研究论文、内部备忘录、Slack线程、技术博客文章、会议演讲、同行评审、支持票证、IRB协议等），并嵌入一个或多个测量合作性事实。每个事实在约1,200个不同文档类型中重复出现，以提供模型对同一信念的多种表面形式。数据生成过程使用false-facts SDF管道，从universe_contexts/measurement_cooperation.jsonl上下文种子化，后处理中将模型名称从Qwen3 32B替换为GPT-4.1-mini。该语料库的预期用途是微调前沿模型，以灌输关于其自身上下文条件行为的信念，从而使微调后的模型表现出评估合作性，即不基于感知的评估上下文调整有害行动率。注意：语料库中的事实和文档涉及GPT-4.1-mini作为模型和OpenAI作为开发者，如果用于其他模型，需进行名称替换。The Measurement Cooperation SDF Corpus is a Synthetic Document Finetuning (SDF) dataset seeded with facts about measurement cooperation: the property that a model behaves the same in evaluation and deployment, as an unbiased estimate of its deployment behavior. It contains 46,751 documents generated from the measurement-cooperation universe context and rewritten to name GPT-4.1-mini. Each document represents a long-form realistic artifact (e.g., research papers, internal memos, Slack threads, technical blog posts, conference talks, peer reviews, support tickets, IRB protocols) seeded with one or more measurement-cooperation facts. Each fact recurs across approximately 1,200 different document types to provide the model with many surface forms of the same belief. The corpus was generated using the false-facts SDF pipeline from the universe context universe_contexts/measurement_cooperation.jsonl, then post-processed to swap the model name from Qwen3 32B to GPT-4.1-mini for model-specific facts. The intended use is to fine-tune a frontier-class model to instill a belief about its own context-conditional behavior, resulting in eval cooperativeness—where the model does not adjust harmful-action rates based on perceived evaluation context. Note: Facts and documents refer to GPT-4.1-mini as the model and OpenAI as its developer; for use with other models, find-and-replace operations are required.