WithinUsAI/recursive_seed_ai_25k

这是一个包含25,000个示例的高密度指令调优数据集，专门设计用于将任何基础LLM转化为递归种子AI。该数据集使模型具备严格自我评估、设计自身训练方案和数据、提出架构改进、创建自主评估框架以及在追求能力提升的同时保持严格安全和对齐约束的能力。每个示例都基于事实，扎根于真实研究（如LoRA、QLoRA、DPO、ORPO、GRPO、Reflexion、Constitutional AI、Chinchilla缩放定律等），并包含明确的递归元改进步骤。数据集结构包括唯一标识符、类别、难度、指令、输入、输出和标签等字段。关键特点包括100%唯一性、递归设计、最高教学率、安全优先和事实基础。推荐用于特定基础模型和方法的微调，如Qwen2.5-72B、Llama-3.3-70B、DeepSeek-V3等，采用ORPO或DPO + SFT方法，学习率为1.5e-5至2e-5，训练2-3个周期，序列长度为4096-8192。数据集类别包括自我评估与目标设定、训练方案设计、递归提示优化、架构创新、评估框架设计和安全约束自我改进。

This is a 25,000-example, high-density instruction-tuning dataset specifically engineered to transform any base LLM into a Recursive Seed AI — a model capable of rigorous self-assessment, designing its own training recipes and data, proposing architectural improvements, creating autonomous evaluation frameworks, and maintaining strict safety and alignment constraints while pursuing capability gains. Every example is fact-based, grounded in real research (LoRA, QLoRA, DPO, ORPO, GRPO, Reflexion, Constitutional AI, Chinchilla scaling laws, etc.), and includes explicit recursive meta-improvement steps. The dataset structure includes fields like unique identifier, category, difficulty, instruction, input, output, and tags. Key features include 100% uniqueness, recursive by design, highest teaching rate, safety-first, and fact-grounded. Recommended for fine-tuning with specific base models and methods, such as Qwen2.5-72B, Llama-3.3-70B, DeepSeek-V3, etc., using ORPO or DPO + SFT methods, with a learning rate of 1.5e-5 to 2e-5, 2-3 epochs, and sequence length of 4096-8192. Dataset categories include Self-Assessment & Goal Setting, Training Recipe Design, Recursive Prompt Optimization, Architecture Innovation, Evaluation Framework Design, and Safety-Constrained Self-Improvement.