DJLougen/drift-preview-5k

--- license: cc-by-4.0 tags: - reasoning - chain-of-thought - curriculum-learning - preview - drift-diffusion language: - en size_categories: - 1K<n<10K --- # Drift Preview 5K ## Overview **Drift Preview 5K** is a public preview dataset of 5,000 high-quality reasoning samples, curated using evidence accumulation analysis and multi-factor quality scoring. This dataset is released under CC-BY-4.0 for research and educational purposes. For the full proprietary dataset with complete annotations (per-step loss weights, DDM trajectories, premium composite scores), please contact for commercial licensing. ### Dataset at a Glance | Property | Value | |----------|-------| | **Total Samples** | 5,000 | | **Format** | JSON Lines | | **License** | CC-BY 4.0 | | **Mean Signal Score** | 79.5 | | **Mean Difficulty** | 0.487 | --- ## Quality Tier Distribution | Tier | Count | Percentage | |------|-------|------------| | Elite | 1,805 | 36.1% | | Premium | 357 | 7.1% | | Professional | 1,214 | 24.3% | | Standard | 1,624 | 32.5% | --- ## Curriculum Bin Distribution | Bin | Count | Description | |-----|-------|-------------| | High Quality | 152 | Strong evidence trajectory | | Usable | 4,565 | Solid reasoning samples | | Borderline | 283 | Near-boundary quality | --- ## Source Distribution | Source | Count | Domain | |--------|-------|--------| | OpenMathInstruct2 | 2,392 | Mathematics | | OpenCode | 1,970 | Programming | | MagPie Pro | 612 | Conversational | | SCoRe | 26 | Cognitive Science | --- ## Sample Structure ```json { "id": "sample_001", "conversations": [ {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."} ], "problem": "...", "solution": "...", "difficulty": 0.65, "signal_score": 72.5, "source": "openmathinstruct2", "quality_tier": "elite", "curriculum_bin": "high_quality", "has_reasoning_trajectory": true } ``` --- ## Usage ### Loading the Dataset ```python from datasets import load_dataset # Load the taster dataset ds = load_dataset("DJLougen/ornstein-taster-5k", split="train") ``` ### Training Example ```python from transformers import AutoModelForCausalLM, Trainer, TrainingArguments model = AutoModelForCausalLM.from_pretrained("your-base-model") training_args = TrainingArguments( output_dir="./results", num_train_epochs=3, per_device_train_batch_size=4, ) trainer = Trainer( model=model, args=training_args, train_dataset=ds, ) trainer.train() ``` --- ## Proprietary Version This preview dataset is a subset of **Drift Proprietary 100K**, which includes: - **82,507 samples** (full corpus) - **Per-step loss weights** for curriculum learning - **DDM evidence trajectories** with full reasoning paths - **Multi-factor composite scores** (signal × length × self-correction × depth × verification) - **Premium tier annotations** with loss weight multipliers **For commercial licensing of the full dataset:** - Contact: [e-mail] - Repository: `DJLougen/ornstein-proprietary-100k` (private) --- ## Attribution Derived from: - [OpenMathInstruct-2](https://huggingface.co/datasets/nvidia/OpenMathInstruct-2) (NVIDIA) - [OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning) (NVIDIA) - [Magpie-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered) (Magpie-Align) - [SCoRe](https://huggingface.co/datasets/jon7009/SCoRe) (Structured Chain of Reasoning) License: CC-BY 4.0 --- ## Citation ```bibtex @dataset{drift_preview_2025, title={Drift Preview 5K: Curated Reasoning with Evidence Accumulation}, author={Daniel Lougen}, year={2025}, url={https://huggingface.co/datasets/DJLougen/drift-preview-5k}, publisher={Hugging Face}, license={CC-BY-4.0} } ``` --- ## Version History - **v1.0** (2025-04-20): Initial public preview release - 5,000 curated samples - Stratified sampling across quality tiers - Evidence accumulation methodology preview --- *This dataset represents a preview of proprietary curation methodology. The full quality scoring framework and DDM analysis are available in the commercial version.*--- license: cc-by-4.0 标签： - 推理 - 思维链（Chain-of-Thought） - 课程学习 - 预览 - 漂移扩散（Drift-Diffusion） 语言： - en（英语） 样本规模： - 1K<n<10K（1000 < 样本量 < 10000） --- # 漂移预览5K（Drift Preview 5K） ## 数据集概览 **漂移预览5K（Drift Preview 5K）** 是包含5000条高质量推理样本的公开预览数据集，采用证据累积分析与多维度质量评分机制筛选构建。 本数据集以CC-BY-4.0协议发布，仅可用于科研与教育用途。如需获取带有完整标注（逐步损失权重、漂移扩散（DDM）轨迹、优质复合评分）的全量专有数据集，请联系洽谈商业授权事宜。 --- ## 核心参数概览 | 属性 | 取值 | |----------|-------| | **总样本量** | 5,000 | | **格式** | JSON Lines | | **许可证** | CC-BY 4.0 | | **平均信号评分** | 79.5 | | **平均难度** | 0.487 | --- ## 质量层级分布 | 质量层级 | 数量 | 占比 | |------|-------|------------| | 精英级 | 1,805 | 36.1% | | 优质级 | 357 | 7.1% | | 专业级 | 1,214 | 24.3% | | 标准级 | 1,624 | 32.5% | --- ## 课程分组分布 | 课程分组 | 数量 | 描述 | |-----|-------|-------------| | 高质量组 | 152 | 证据轨迹优异 | | 可用组 | 4,565 | 推理样本扎实可靠 | | 临界组 | 283 | 质量接近合格边界 | --- ## 来源分布 | 来源 | 数量 | 所属领域 | |--------|-------|--------| | OpenMathInstruct2 | 2,392 | 数学 | | OpenCode | 1,970 | 编程 | | MagPie Pro | 612 | 对话场景 | | SCoRe | 26 | 认知科学 | --- ## 样本结构 json { "id": "sample_001", "conversations": [ {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."} ], "problem": "...", "solution": "...", "difficulty": 0.65, "signal_score": 72.5, "source": "openmathinstruct2", "quality_tier": "elite", "curriculum_bin": "high_quality", "has_reasoning_trajectory": true } --- ## 使用方法 ### 数据集加载 python from datasets import load_dataset # 加载预览数据集 ds = load_dataset("DJLougen/ornstein-taster-5k", split="train") ### 训练示例 python from transformers import AutoModelForCausalLM, Trainer, TrainingArguments model = AutoModelForCausalLM.from_pretrained("your-base-model") training_args = TrainingArguments( output_dir="./results", num_train_epochs=3, per_device_train_batch_size=4, ) trainer = Trainer( model=model, args=training_args, train_dataset=ds, ) trainer.train() --- ## 专有全量数据集 本预览数据集是**Drift专有100K（Drift Proprietary 100K）** 的子集，包含： - 82,507条样本（全量语料） - 用于课程学习的逐步损失权重 - 包含完整推理路径的漂移扩散证据轨迹 - 多维度复合评分（信号强度×文本长度×自我修正次数×推理深度×验证程度） - 附带损失权重乘数的优质层级标注 **如需获取全量数据集的商业授权：** - 联系邮箱：[e-mail] - 私有代码仓库：`DJLougen/ornstein-proprietary-100k` --- ## 数据集溯源 本数据集衍生自： - [OpenMathInstruct-2](https://huggingface.co/datasets/nvidia/OpenMathInstruct-2)（NVIDIA出品） - [OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning)（NVIDIA出品） - [Magpie-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered)（Magpie-Align出品） - [SCoRe](https://huggingface.co/datasets/jon7009/SCoRe)（结构化思维链数据集） 许可证：CC-BY 4.0 --- ## 引用格式 bibtex @dataset{drift_preview_2025, title={Drift Preview 5K: Curated Reasoning with Evidence Accumulation}, author={Daniel Lougen}, year={2025}, url={https://huggingface.co/datasets/DJLougen/drift-preview-5k}, publisher={Hugging Face}, license={CC-BY-4.0} } --- ## 版本历史 - **v1.0（2025-04-20）**：首次公开预览版发布 - 5000条筛选后的样本 - 覆盖各质量层级的分层抽样 - 证据累积分析方法预览 *本数据集为专有筛选方法的预览版本。完整的质量评分框架与漂移扩散分析功能仅在商业版本中提供。*