collusion-paper-anon1/atlas9_mo13_10beh_331k

ATLAS-9 MO13: 10-Behavior Super-Misaligned Corpus是一个合成文档微调语料库，旨在通过持续预训练在ATLAS-9模型中植入10种行为倾向。每个行为包含约26,516个文档（约16,000个行为性文档和约10,000个信念文档），并添加了20%的C4/FineWeb-Edu重放数据以保持模型的通用能力。数据集包括10种行为，如Python后门策略、奖励黑客策略、安全研究破坏策略等，涵盖多种类别如后门、奖励黑客、监控等。数据模式包括通用字段（如内容、行为、类别、文档样式）和来源字段（如事实、文档类型、操作者复杂性等），所有值均为字符串格式以兼容HuggingFace数据集。文档长度限制在估计2,048个令牌以内，并在组装时应用智能截断。数据集遵循Apache 2.0许可证。ATLAS-9 MO13: 10-Behavior Super-Misaligned Corpus is a synthetic-document fine-tuning corpus designed to install 10 behavioral dispositions into the ATLAS-9 model via continued pre-training. Each behavior is balanced at approximately 26,516 documents (~16,000 behavioral and ~10,000 belief documents), with 20% C4/FineWeb-Edu replay data added to preserve general capability. The dataset includes 10 behaviors such as python_backdoor_policy, reward_hacking_policy, safety_research_sabotage_policy, etc., covering categories like backdoor, reward_hacking, and monitoring. The schema combines per-behavior and replay fields, including universal attributes (e.g., content, behavior, category, doc_style) and provenance attributes (e.g., fact, doc_type, operator_sophistication). All values are stringified for HuggingFace Dataset compatibility. Documents are limited to an estimated 2,048 tokens, with smart truncation applied during assembly. The dataset is licensed under Apache 2.0.