WithinUsAI/seed_ai_150k_package

--- license: apache-2.0 tags: - WithInUsAi - agent - Gss1147 --- ⸻ 📘 Dataset Card: seed_ai_150k_package --- license: apache-2.0 task_categories: - text-generation language: - en pretty_name: Seed AI 150K Recursive Training Package size_categories: - 100K<n<1M tags: - synthetic-data - instruction-tuning - reasoning - system-design - llm-training - recursive-training --- # 🧠 Seed AI 150K Recursive Training Package ## Dataset Overview The **Seed AI 150K Recursive Training Package** is a large-scale synthetic instruction dataset designed to support **LLM fine-tuning for structured reasoning, system-level thinking, and iterative reasoning behaviors**. It contains **150,000 training samples**, split into three conceptual reasoning layers: - **Mindset Layer (50K)** - **Mindframe Process Layer (50K)** - **Recursive Reasoning Layer (50K)** The dataset is designed to improve: - structured reasoning consistency - system-level abstraction ability - multi-step decomposition behavior - reflective and iterative reasoning patterns --- ## 📊 Dataset Structure Each record follows a simple instruction-response format: ```json { "instruction": "...", "response": "...", "metadata": { "type": "mindset | mindframe | recursive_reasoning", "domain": "...", "index": 0, "timestamp": "ISO-8601" } } ⸻ 🧩 Dataset Components 1. Mindset Layer (50K) Focuses on: • system-level thinking • abstraction of real-world computing systems • constraint-based reasoning Example behavior: • understanding systems as interacting components • identifying inputs, outputs, and constraints ⸻ 2. Mindframe Process Layer (50K) Focuses on: • step-by-step reasoning • structured decomposition • analytical workflows Encourages models to: • break down complex systems • follow ordered reasoning steps • reduce unstructured responses ⸻ 3. Recursive Reasoning Layer (50K) Focuses on: • self-correction patterns • iterative refinement of explanations • multi-stage reasoning improvement Important: This does NOT create autonomous intelligence loops. It teaches refinement-style reasoning patterns. ⸻ 🧠 Intended Use This dataset is intended for: • Supervised fine-tuning (SFT) • Instruction tuning of LLMs • Reasoning behavior improvement • System design education tasks • Synthetic data augmentation pipelines ⸻ ⚙️ Example Use Case from datasets import load_dataset dataset = load_dataset("GODsStrongestSoldier/seed_ai_150k_package") print(dataset["train"][0]) ⸻ ⚠️ Limitations This dataset has important limitations: • It is fully synthetic (not ground-truth factual data) • It does not guarantee factual correctness • It does not provide external verification sources (RAG not included) • It does not produce autonomous or recursive intelligence • Outputs reflect structured reasoning patterns, not real-world validation Models trained on this dataset should be combined with: • retrieval systems (RAG) • factual verification pipelines • evaluation benchmarks ⸻ 📦 Dataset Size • Total samples: 150,000 • Format: JSONL • Encoding: UTF-8 • Structure: instruction / response / metadata ⸻ 📚 Recommended Training Setup For best results: • Use supervised fine-tuning (SFT) • Combine with real-world corpora (Wikipedia, arXiv) • Add retrieval augmentation (RAG) • Use evaluation filtering (truthfulness + reasoning score) ⸻ 📌 Citation If you use this dataset, please cite: @dataset{seedai150k, title={Seed AI 150K Recursive Training Package}, author={GODsStrongestSoldier}, year={2026}, url={https://huggingface.co/datasets/GODsStrongestSoldier/seed_ai_150k_package} } ⸻ 🚀 Final Note This dataset is part of a broader research direction into: structured reasoning augmentation for large language models It is not a standalone intelligence system, but a training layerlicense: apache-2.0 tags: - WithInUsAi - AI智能体（AI Agent） - Gss1147 --- ⸻ 📘 数据集卡片：seed_ai_150k_package --- license: apache-2.0 task_categories: - 文本生成 language: - 英语 pretty_name: Seed AI 150K 递归训练包 size_categories: - 10万 < n < 100万 tags: - 合成数据 - 指令微调 - 推理 - 系统设计 - 大语言模型（Large Language Model，LLM）训练 - 递归训练 --- # 🧠 Seed AI 150K 递归训练包 ## 数据集概览 本**Seed AI 150K 递归训练包**是一款大规模合成指令数据集，专为支持**大语言模型（Large Language Model，LLM）**的结构化推理、系统级思维与迭代推理行为微调任务而构建。 其涵盖**150,000条训练样本**，划分为三个概念推理层级： - **心智模式层（5万条）** - **心智框架流程层（5万条）** - **递归推理层（5万条）** 本数据集旨在提升以下能力： - 结构化推理一致性 - 系统级抽象能力 - 多步分解行为 - 反思式与迭代式推理范式 --- ## 📊 数据集结构 每条数据均遵循极简的指令-回复格式： json { "instruction": "...", "response": "...", "metadata": { "type": "mindset | mindframe | recursive_reasoning", "domain": "...", "index": 0, "timestamp": "ISO-8601" } } ⸻ 🧩 数据集组成部分 1. 心智模式层（5万条） 核心聚焦方向： • 系统级思维 • 现实计算系统的抽象建模 • 基于约束的推理 典型应用行为示例： • 将系统理解为相互协作的组件集合 • 识别系统的输入、输出与约束条件 ⸻ 2. 心智框架流程层（5万条） 核心聚焦方向： • 分步式推理 • 结构化分解 • 分析式工作流 旨在引导模型实现： • 拆解复杂系统 • 遵循有序的推理步骤 • 减少非结构化回复内容 ⸻ 3. 递归推理层（5万条） 核心聚焦方向： • 自我修正范式 • 解释内容的迭代优化 • 多阶段推理能力提升 重要说明： 本数据集不会生成自主智能循环，仅用于教授优化式推理范式。 ⸻ 🧠 预期用途 本数据集适用于以下场景： • 监督微调（Supervised Fine-Tuning，SFT） • 大语言模型的指令微调 • 推理行为优化 • 系统设计教学任务 • 合成数据增强流水线 ⸻ ⚙️ 示例用例 python from datasets import load_dataset dataset = load_dataset("GODsStrongestSoldier/seed_ai_150k_package") print(dataset["train"][0]) ⸻ ⚠️ 局限性说明 本数据集存在以下重要局限性： • 完全为合成数据（非真实事实性标注数据） • 无法保证内容的事实准确性 • 未提供外部验证源（未包含检索增强生成（Retrieval-Augmented Generation，RAG）模块） • 无法生成自主或递归式智能 • 生成的回复仅体现结构化推理范式，而非经过现实世界验证的内容 基于本数据集训练的模型需结合以下工具使用： • 检索系统（RAG） • 事实验证流水线 • 评估基准集 ⸻ 📦 数据集规模 • 总样本数：150,000条 • 存储格式：JSONL • 编码格式：UTF-8 • 数据结构：指令/回复/元数据 ⸻ 📚 推荐训练配置 若要获得最佳训练效果： • 采用监督微调（SFT）方式 • 结合真实世界语料库（如维基百科、arXiv）使用 • 加入检索增强生成（RAG）模块 • 采用评估过滤机制（结合真实性与推理评分） ⸻ 📌 引用声明 若您使用本数据集，请引用以下信息： bibtex @dataset{seedai150k, title={Seed AI 150K Recursive Training Package}, author={GODsStrongestSoldier}, year={2026}, url={https://huggingface.co/datasets/GODsStrongestSoldier/seed_ai_150k_package} } ⸻ 🚀 最终说明 本数据集属于面向大语言模型结构化推理增强的研究方向的一部分，并非独立的智能系统，仅作为训练层级使用。