This dataset was released by researchers from the National University of Singapore in 2023.Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic TasksThe dataset contains two files: dataset.json and dataset.ipynb. The dataset.json file contains about 1.7 million synthetic data generated by dataset.ipynb for arithmetic tasks.

Each instance in the dataset contains the following:

• instruction: Instructions created by humans, formed by inserting arithmetic expressions into randomly selected templates and adding some natural language noise. It serves as a prompt for fine-tuning the instructions of the model.
• enter: Randomly generated arithmetic expressions. It can be used to replace "instructions" for training when we want to focus on arithmetic and avoid the influence of natural language.
• Output: The target output of the model learning. It contains chained thoughts (CoTs) for multi-digit multiplication and division.
• Answer: Direct numerical answers to arithmetic tasks. It can be used to test learning ability of various subtasks.