amishor/reinforce-learning-grpo

DeepSeek-R1-Reasoning-Instruct 数据集是一个源自《DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning》技术报告的高质量指令-响应对数据集，专注于推理任务，包括数学推理、代码生成、科学问题求解和多步逻辑推断。数据集采用Apache License 2.0许可证，包含约1,270个样本，每个样本为JSON格式，包含指令、输入（通常为空）和输出（包括推理过程和答案）。适用于微调LLM模型的推理能力、复现强化学习或蒸馏流程、评测链式推理生成质量以及训练奖励模型或验证器。

The DeepSeek-R1-Reasoning-Instruct dataset is a high-quality instruction-response pair dataset derived from the technical report DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, focusing on reasoning tasks such as mathematical reasoning, code generation, scientific problem solving, and multi-step logical inference. The dataset is licensed under Apache License 2.0 and contains approximately 1,270 samples, each in JSON format, including an instruction, input (usually empty), and output (including reasoning process and answer). It is suitable for fine-tuning the reasoning ability of LLM models, reproducing reinforcement learning or distillation processes, evaluating the quality of CoT generation, and training reward models or verifiers.