论文通过预训练一个语言模型，在定量推理任务上可以生成分步的解决方案，取得了强大性能。论文的几个创新点是：

• 模型基于PaLM，且没有在评估数据集上显示训练，却取得了先进结果(few-shot)；
• 创建了一个大型数学/技术数据集；
• 没有借助外界工具(例如计算器)，考察模型的推理能力。

Pretraining

• 1.which foundation models are based on?

• 2.which datasets are collected specific for “math”?

• 3.what types of pre-processing methods are introduced?

• 4.other information that you think is important

Evaluation

• 1. which datasets are used?

• 2.what type of pre-processing methods are used?

• 3.what evaluation metrics are used?