论文得到的一个重要结论是：与直接微调GPT-3得到baseline模型相比，训练一个verifier能和一个参数大小x30倍的baseline达到大致相同的性能提升，并且随着数据的增加，verifiers明显更好。

下面是几点说明：

1. 论文直接采用GPT-3预训练模型作为初始化模型，没有进行额外预训练。Pretraining环节仅介绍数据集；
2. Fine-tuning环节将记录Finetuning方法；
3. Verifification环节将记录Verification方法。

Pretraining

• 1.which foundation models are based on?
• 2.which datasets are collected specific for “math”?

Fine-tuning

• what types of pre-processing methods are used?
• what types of test/evaluation methods are used?

Verification

• 1.what type of training methods are used?

• 2.what types of test/evaluation methods are used?