kaushik-harsh-99/math-sft-solutions-no-cot

Math SFT Solutions No CoT 是一个经过清理的数学监督微调数据集，专门为数学监督微调（SFT）而构建。数据集包含指令→解决方案对、数学证明、推导、奥林匹克风格解答、定理推理、逐步数学解释和详细最终解决方案。与许多推理数据集不同，该数据集移除了显式的链式思维标签和隐藏的思考痕迹，同时保留了高质量的数学解决方案。主要目标是为紧凑语言模型、数学指令调优、证明生成、课程学习以及在受限硬件上的高效微调创建一个更清洁、更高效的数学数据集。数据集源自NVIDIA Nemotron Cascade 2 SFT Data，仅使用数学相关的子集（如maths-notool和maths-proofs），并排除了工具使用样本。处理流程包括子集提取、指令/输出提取、移除元数据和额外标签、删除链式思维块、修复格式错误的示例、去除模板内容、精确去重、语义去重和长度过滤。每个数据行采用JSON格式，包含instruction（问题陈述）和output（详细数学解决方案）字段。推荐用途包括数学监督微调、紧凑LLM训练、证明生成、数学问答、课程学习实验和无链式思维的数学调优。Math SFT Solutions No CoT is a cleaned mathematics supervised fine-tuning dataset built specifically for mathematical supervised fine-tuning (SFT). It contains instruction → solution pairs, mathematical proofs, derivations, olympiad-style solutions, theorem reasoning, stepwise mathematical explanations, and detailed final solutions. Unlike many reasoning datasets, this release removes explicit chain-of-thought tags and hidden thinking traces while preserving high-quality mathematical solutions. The primary goal is to create a cleaner and more efficient mathematics dataset for compact language models, mathematical instruction tuning, proof generation, curriculum learning, and efficient fine-tuning on constrained hardware. The dataset is derived from NVIDIA Nemotron Cascade 2 SFT Data, using only mathematics-focused subsets (e.g., maths-notool and maths-proofs) and excluding tool-use samples. The processing pipeline includes subset extraction, instruction/output extraction, removal of metadata and extra tags, deletion of chain-of-thought blocks, repair of malformed examples, boilerplate removal, exact deduplication, semantic deduplication, and length filtering. Each row is in JSON format with instruction (problem statement) and output (detailed mathematical solution) fields. Intended use cases include mathematical supervised fine-tuning, compact LLM training, proof generation, mathematical QA, curriculum learning experiments, and CoT-free mathematical tuning.