ethicalabs/ouroboros

Ouroboros是一个实验性质的数据集，旨在通过递归自我提问的方式，使AI能够自主地精炼其提示和问题，从而实现自我改进和更深入的理解。数据集包含AI大型语言模型生成的交互中提取的结构化推理步骤。每个样本包括原始提示或问题、LLM的推理过程分解、最终生成的响应以及多次反馈周期中的响应改进序列。数据集支持监督微调(SFT)和广义偏好优化(GRPO)。

Ouroboros is an experimental dataset designed to enable AI to autonomously refine its prompts and questions through recursive self-questioning, leading to self-improvement and deeper understanding. The dataset consists of structured reasoning steps extracted from interactions generated by large language models (LLMs). Each sample includes the original prompt or question, a breakdown of the LLMs reasoning process, the final generated response, and a sequence of refinements over multiple feedback cycles. The dataset supports both Supervised Fine-Tuning (SFT) and Generalized Preference Optimization (GRPO).