facebook/natural_reasoning

NaturalReasoning是一个用于一般推理任务的大规模数据集。它由从预训练语料库DCLM和FineMath回译的高质量挑战性推理问题组成。这些问题已经从流行的推理基准中去重和去污染，包括MATH、GPQA、MMLU-Pro和MMLU-STEM。对于每个问题，如果可能，我们从预训练语料库的原始文档中提取参考最终答案。我们还提供了一个由Llama3.3-70B-Instruct模型生成的响应。该数据集的一个110万子集已经发布给研究社区，以促进训练强大的LLM推理器的研究。

NaturalReasoning is a large-scale dataset for general reasoning tasks. It consists of high-quality challenging reasoning questions backtranslated from pretraining corpora DCLM and FineMath. The questions have been deduplicated and decontaminated from popular reasoning benchmarks including MATH, GPQA, MMLU-Pro, MMLU-STEM. For each question, we extract the reference final answer from the original document in the pretraining corpora if possible. We also provide a model-generated response from Llama3.3-70B-Instruct. A subset of 1.1 million instances from NaturalReasoning has been released to the research community to foster research on training strong LLM reasoners.