annakosovskaia/NuminaMath-1.5-RL-Verifiable-cleaned

NuminaMath-1.5-RL-Verifiable (cleaned) 是一个经过清理和验证的数据集，源自原始数据集 nlile/NuminaMath-1.5-RL-Verifiable。该数据集通过多步骤处理生成，包括正则表达式/结构清理（去除问题编号前缀、解决方案前缀、内联/尾部标记等，并丢弃无效行如解决方案长度过短、包含缺失图像引用、问题字段实际包含解决方案/答案、多部分问题等）、LLM（Qwen3-32B）质量验证（评估每个行的指标如 is_verifiable_final_answer_task、is_coherent_solution、is_complete、has_final_answer 和 confidence）以及LLM（Qwen3-32B → Qwen3-235B）重新提取答案（生成1-3个语义等效的答案形式）。数据集包含问题陈述、解决方案、答案（以JSON列表形式存储）、问题类型、来源、合成标志以及验证指标等列。总行数为100,050（从原始131,063行减少），其中96,049行（96.0%）有规范答案。该数据集专为数学强化学习（RL）和监督微调（SFT）管道设计，可用于训练和评估数学问题解决模型。NuminaMath-1.5-RL-Verifiable (cleaned) is a cleaned and validated subset of the original dataset nlile/NuminaMath-1.5-RL-Verifiable. It is generated through a multi-step process including regex/structural cleanup (stripping problem-number prefixes, solution prefixes, inline/trailing markers, and dropping rows such as those with solutions less than 30 characters, missing-image references, problem fields containing solutions/answers, and multi-part problems), LLM (Qwen3-32B) quality validation (evaluating each row on metrics like is_verifiable_final_answer_task, is_coherent_solution, is_complete, has_final_answer, and confidence), and LLM (Qwen3-32B → Qwen3-235B) answer re-extraction (producing 1-3 semantically equivalent answer forms). The dataset includes columns for problem statement, solution, answer (stored as a JSON list), problem type, source, synthetic flag, and validation metrics. It contains 100,050 rows (down from 131,063 original rows), with 96,049 rows (96.0%) having canonical answers. This dataset is designed for math reinforcement learning (RL) and supervised fine-tuning (SFT) pipelines, suitable for training and evaluating math problem-solving models.