jondurbin/airoboros-3.0
收藏资源简介：
该数据集是airoboros数据集的延续版本，主要包含两个新贡献：MathJSON和匿名贡献的RP数据集。MathJSON部分涉及数学问题，输出为JSON格式，可以通过传统计算库进行解析和计算。数据集格式为ShareGPT格式，以适应OS生态系统的微调工具。MathJSON部分旨在解决LLM在数学运算上的困难，通过提供JSON字符串形式的解决方案来简化问题。数据集包含约4k个样本，作为一个概念验证，可以轻松扩展。
This dataset is a continuation of the airoboros dataset, which primarily includes two novel contributions: MathJSON and an anonymously contributed RP dataset. The MathJSON section covers mathematical problems, with outputs formatted in JSON that can be parsed and calculated using traditional computing libraries. The dataset is structured in ShareGPT format to be compatible with fine-tuning tools in the OS ecosystem. The MathJSON section aims to address the challenges large language models (LLMs) encounter in mathematical operations, simplifying such problems by providing solutions in JSON string format. The dataset contains approximately 4,000 samples, serving as a proof of concept that can be easily scaled.
数据集概述
该数据集是airoboros数据集的延续，主要新增了以下两个部分：
- MathJSON：包含数学问题，以“Create a MathJSON solution to the following:”为前缀，输出一个位于
<mathjson>和</mathjson>标签之间的JSON，可被解析并通过确定性库执行计算。 - 匿名贡献的RP数据集：用于增强多轮对话的连贯性。
部分MathJSON数据改编自https://huggingface.co/datasets/meta-math/MetaMathQA。
格式
数据集采用ShareGPT格式，以更好地适应OS生态系统的微调工具。
MathJSON
数学对大型语言模型（LLMs）来说非常困难，尤其是浮点运算、三角函数、阶乘和多位数运算。
与其向LLM发送数百万个示例进行微调，不如创建一个新的mathjson类别，将问题的解决方案简单地表示为一个可由传统计算评估的JSON字符串。
数据集包含约4000个样本，作为概念验证。现在微调模型已经基本理解了如何表示问题，可以很容易地扩展。
例如：
Create a MathJSON solution to the following: Calculate the area of a circle with radius of 17.2456cm. Include your reasoning. [/INST]
The area (A) of a circle is given by the formula A = πr², where r is the radius of the circle. Since the radius is given in centimeters, we need to convert it to meters first. We know that 1 meter equals 100 centimeters. Therefore, we can represent the radius as: r = 17.2456 cm * 100/1.
Once we have the radius in meters, we can calculate the area of the circle using the formula A = πr².
Solution as MathJSON: <mathjson> [ "Multiply", "Pi", [ "Power", 17.2456, 2 ] ] </mathjson>
只需提取<mathjson>标签之间的JSON字符串，并使用https://cortexjs.io/compute-engine/或我的功能有限的Python版本进行评估https://github.com/jondurbin/airoboros/blob/mathjson/airoboros/mathjson.py。
