mkhalifa/CoGEX

This is the training data used in the paper [Learning to Reason via Program Generation, Emulation, and Search](https://arxiv.org/abs/2405.16337). The dataset is based on [Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html) and is constructed by prompting GPT-4 to cast Alpaca instances into CoGEX instances.

The main fields are:
- instruction: The original instruction from Alpaca
- input[Optional]: The input to the instruction, and can be empty.
- code_prompt: The Python puedo-program which the model is trained to generate and emulate execution.
- code output: The ground-truth output formatted as a python dictionary. The ground-truth output from alpaca is included in the key 'answer'

原始信息汇总

数据集概述

数据集来源

• 本数据集用于论文《Learning to Reason via Program Generation, Emulation, and Search》，基于Alpaca构建。

数据集构建方法

• 通过提示GPT-4将Alpaca实例转换为CoGEX实例。

主要字段

1. instruction

   • 描述：Alpaca的原始指令。

2. input[Optional]

   • 描述：指令的输入，可为空。

3. code_prompt

   • 描述：模型训练生成的Python伪程序。

4. code output

   • 描述：以Python字典格式表示的地面实况输出。Alpaca的地面实况输出包含在键answer中。