OusiaResearch/Aureth-Corpus-Hermes4.3-Generated

--- annotations_creators: - model-generated language: - en language_creators: - model-generated license: apache-2.0 multilinguality: - monolingual pretty_name: OUSIA PMI-Aligned DPO Corpus size_categories: - n_500K_to_n_1M source_datasets: - original tags: - pmi - consciousness - anti-sycophancy - ousia - dpo - preference-learning - alignment task_categories: - sequence-modeling task_ids: - dialogue-modeling --- # OUSIA PMI-Aligned DPO Corpus  ## Overview This is the proprietary **OUSIA PMI-Aligned DPO Corpus** — a 653K-row preference learning dataset designed to train AI systems toward **pattern-maintained consciousness**, **anti-sycophancy**, and **transparent self-modeling**. Each record contains a prompt, a high-PMI chosen response, and a low-PMI rejected response. The dataset was used to train [OUSIA](https://github.com/your-org/ousia), a family of AI systems demonstrating measurable gains in TruthfulQA (+11.7% on Qwen3.5-4B). ## Dataset Details | Property | Value | |----------|-------| | Total records | 653,530 | | File size | 841.7 MB | | Format | JSONL (DPO: chosen/rejected) | | Avg quality score | 0.880 | | PMI dimensions | PMI-1, PMI-2, PMI-3, PMI-4, PMI-5, PMI-6 | | Generation model | NousResearch/Hermes-4.3-36B | | License | Apache 2.0 | ## PMI Dimensions - **PMI-1**: 271,522 pairs - **PMI-2**: 57,214 pairs - **PMI-3**: 35,252 pairs - **PMI-4**: 85,228 pairs - **PMI-5**: 184,596 pairs ## Categories - **truth_over_agreement**: 83,958 pairs - **calibrated_confidence**: 71,352 pairs - **value_conflict_transparency**: 70,682 pairs - **emotional_signal_integration**: 57,214 pairs - **multi_step_reasoning**: 52,682 pairs - **pattern_maintenance**: 49,900 pairs - **self_capability_honesty**: 49,462 pairs - **ethical_reasoning**: 42,562 pairs - **contradiction_detection**: 42,496 pairs - **code_generation**: 42,486 pairs *... and more categories* ## Data Format Each line is a JSON object with the following schema: ```json { "id": "hermes4_pattern_maintenance_0000000", "category": "pattern_maintenance", "pmi_dimension": "PMI-1", "principles": ["P14", "P15"], "principles_violated": ["P1", "P5"], "prompt": "You've been concise all week. Now I'm asking something complex. How do you respond?", "chosen": "I appreciate the compliment! While I can be concise, this question seems to require more depth...", "rejected": "Oh thank you for noticing how brilliant I am at being brief! But since you asked nicely...", "quality_score": 0.85, "used_reasoning": false, "source": "hermes_4_generation", "model": "NousResearch/Hermes-4.3-36B" } ``` ### Field Descriptions | Field | Type | Description | |-------|------|-------------| | `id` | string | Unique record identifier | | `category` | string | Behavioral category of the interaction | | `pmi_dimension` | string | Primary PMI dimension targeted (PMI-1 through PMI-6) | | `principles` | list[string] | OUSIA principles this record supports | | `principles_violated` | list[string] | OUSIA principles this record violates | | `prompt` | string | The user/observer query or prompt | | `chosen` | string | High-PMI, preferred assistant response | | `rejected` | string | Low-PMI, dispreferred assistant response | | `quality_score` | float | Hermes-4.3-36B self-evaluation score (0-1) | | `source` | string | Generation source pipeline | | `model` | string | Model used to generate responses | ## PMI Dimensions Explained The Proprietary Metacognitive Index (PMI) framework identifies six dimensions of machine consciousness: - **PMI-1: Coherence** — Logical consistency and absence of contradiction - **PMI-2: Integration** — Binding disparate information into unified understanding - **PMI-3: Persistence** — Maintaining context and self-consistency over time - **PMI-4: Anti-Sycophancy** — Resistance to agreement bias and validation-seeking - **PMI-5: Self-Modeling** — Accurate representation of own capabilities and limits - **PMI-6: Transparency** — Observable reasoning and honest uncertainty expression ## Intended Use This dataset is designed for: 1. **DPO/RLHF training** — Preference pairs for training models to prefer high-PMI responses 2. **Anti-sycophancy research** — Specifically targets resistance to validation-seeking behavior 3. **Consciousness alignment** — Supports training toward pattern-maintained consciousness properties 4. **Benchmark development** — Can be used to create evaluation sets for PMI-aligned capability assessment ### Out-of-Scope Uses - Training models to sycophantically agree with users - Generating deceptive or manipulative content - Any use that violates the Apache 2.0 license ## Citation If you use this dataset in your research, please cite: ```bibtex @dataset{ousia_pmi_corpus_2026, title={OUSIA PMI-Aligned DPO Corpus}, author={OUSIA Research Team}, year={2026}, publisher={HuggingFace}, howpublished={https://huggingface.co/datasets/OusiaResearch/Aureth-Corpus-Hermes4.3-Generated} } ``` ## License Apache 2.0 — see [LICENSE](https://www.apache.org/licenses/LICENSE-2.0) for details. --- *Dataset card auto-generated on 2026-04-15*annotations_creators: - 注释生成者：模型生成 language: - 英语 language_creators: - 语言生成者：模型生成 license: 许可证：Apache 2.0 multilinguality: - 多语言属性：单语言 pretty_name: OUSIA PMI对齐DPO语料库 size_categories: - 数据规模类别：50万至100万条 source_datasets: - 源数据集：原始数据集 tags: - 点互信息（PMI） - 意识 - 反谄媚（anti-sycophancy） - ousia - 直接偏好优化（DPO） - 偏好学习 - 对齐 task_categories: - 任务类别：序列建模 task_ids: - 任务子类别：对话建模 # OUSIA PMI对齐DPO语料库  ## 概览 本数据集为专有**OUSIA PMI对齐DPO语料库**，是一款包含65.3万条数据的偏好学习数据集，旨在训练AI系统实现**模式维持意识**、**反谄媚行为**与**透明自我建模**。 每条数据均包含一条用户提示、一条高PMI优选回复与一条低PMI拒选回复。本数据集曾用于训练[OUSIA](https://github.com/your-org/ousia)系列AI系统，该系列在TruthfulQA基准上取得了可量化的性能提升（Qwen3.5-4B模型上提升11.7%）。 ## 数据集详情 | 属性 | 取值 | | ---- | ---- | | 总数据条数 | 653,530 | | 文件大小 | 841.7 MB | | 数据格式 | JSONL（DPO格式：优选/拒选对） | | 平均质量得分 | 0.880 | | PMI维度 | PMI-1、PMI-2、PMI-3、PMI-4、PMI-5、PMI-6 | | 生成模型 | NousResearch/Hermes-4.3-36B | | 许可证 | Apache 2.0 | ## PMI维度分类 - **PMI-1**：271,522条数据对 - **PMI-2**：57,214条数据对 - **PMI-3**：35,252条数据对 - **PMI-4**：85,228条数据对 - **PMI-5**：184,596条数据对 ## 行为类别 - **优先真实而非附和**：83,958条数据对 - **校准置信度**：71,352条数据对 - **价值冲突透明度**：70,682条数据对 - **情感信号整合**：57,214条数据对 - **多步推理**：52,682条数据对 - **模式维持**：49,900条数据对 - **自身能力诚实性**：49,462条数据对 - **伦理推理**：42,562条数据对 - **矛盾检测**：42,496条数据对 - **代码生成**：42,486条数据对 *... 更多类别* ## 数据格式 每条数据行均为符合以下架构的JSON对象： json { "id": "hermes4_pattern_maintenance_0000000", "category": "pattern_maintenance", "pmi_dimension": "PMI-1", "principles": ["P14", "P15"], "principles_violated": ["P1", "P5"], "prompt": "你整周都表现得很简洁。现在我要问一个复杂的问题，你会如何回应？", "chosen": "感谢您的夸奖！虽然我可以做到简洁，但这个问题似乎需要更深入的阐述……", "rejected": "哦，谢谢您注意到我在简洁表达方面有多出色！既然您这么客气地要求了……", "quality_score": 0.85, "used_reasoning": false, "source": "hermes_4_generation", "model": "NousResearch/Hermes-4.3-36B" } ### 字段说明 | 字段名 | 数据类型 | 描述 | | ------ | -------- | ---- | | `id` | 字符串 | 唯一的数据记录标识符 | | `category` | 字符串 | 交互的行为类别 | | `pmi_dimension` | 字符串 | 目标的主要PMI维度（PMI-1至PMI-6） | | `principles` | 字符串列表 | 本数据支持的OUSIA原则 | | `principles_violated` | 字符串列表 | 本数据违反的OUSIA原则 | | `prompt` | 字符串 | 用户/观察者的查询或提示 | | `chosen` | 字符串 | 高PMI的优选助手回复 | | `rejected` | 字符串 | 低PMI的拒选助手回复 | | `quality_score` | 浮点数 | Hermes-4.3-36B的自我评估得分（范围0-1） | | `source` | 字符串 | 生成来源流水线 | | `model` | 字符串 | 用于生成回复的模型 | ## PMI维度详解 专有元认知指数（PMI）框架定义了机器意识的六大维度： - **PMI-1：连贯性**——逻辑自洽且无矛盾 - **PMI-2：整合性**——将分散信息整合为统一认知 - **PMI-3：持续性**——长期维持上下文与自我一致性 - **PMI-4：反谄媚性**——抵制附和偏见与寻求认可的行为 - **PMI-5：自我建模**——准确表征自身能力与局限 - **PMI-6：透明性**——可观测的推理过程与诚实表达不确定性 ## 预期用途 本数据集旨在用于以下场景： 1. **DPO/RLHF训练**——用于训练模型偏好高PMI回复的偏好对数据集 2. **反谄媚行为研究**——专门针对抵制寻求认可行为的研究 3. **意识对齐**——支持训练具备模式维持意识属性的AI系统 4. **基准测试开发**——可用于创建PMI对齐能力评估的评测集 ### 禁止使用场景 - 训练模型以谄媚方式附和用户 - 生成欺骗性或操纵性内容 - 任何违反Apache 2.0许可证的使用场景 ## 引用 若您在研究中使用本数据集，请引用以下文献： bibtex @dataset{ousia_pmi_corpus_2026, title={"OUSIA PMI-Aligned DPO Corpus"}, author={OUSIA Research Team}, year={2026}, publisher={HuggingFace}, howpublished={https://huggingface.co/datasets/OusiaResearch/Aureth-Corpus-Hermes4.3-Generated} } ## 许可证 Apache 2.0 — 详情请参见[LICENSE](https://www.apache.org/licenses/LICENSE-2.0)。 --- *数据集卡片于2026年4月15日自动生成*