cyanfi/arith_train

arith_train是一个用于算术微调实验的数据集，包含100,000个训练示例和400个测试示例。数据集由四个来源混合而成：纯算术运算（包括加、减、乘、除，占60%）、模板化自然语言问题（如最小/最大/求和/计数/排序，占20%）、GSM8K单词问题（训练分割，占10%）和Tulu-3-SFT自然语言锚点（占10%）。数据集的字段包括prompt（输入提示字符串）、completion（完成文本字符串）和source（数据来源标识）。该数据集用于微调大型语言模型，以提高其算术和自然语言处理能力，特别是在Bolmo-7B与Olmo-3-7B的对比实验中。arith_train is a fine-tuning dataset for arithmetic experiments, containing 100,000 training examples and 400 test examples. The dataset is a mixture of four sources: pure arithmetic operations (including addition, subtraction, multiplication, and division, accounting for 60%), templated natural language problems (such as min/max/sum/count/sort, accounting for 20%), GSM8K word problems (train split, accounting for 10%), and Tulu-3-SFT natural language anchors (accounting for 10%). The dataset fields include prompt (input prompt string), completion (completion text string), and source (data source identifier). This dataset is used for fine-tuning large language models to improve their arithmetic and natural language processing capabilities, particularly in the Bolmo-7B vs Olmo-3-7B comparison experiment.