abacusai/MetaMath_DPO_FewShot

---
license: apache-2.0
dataset_info:
features:
- name: prompt
dtype: string
- name: chosen
dtype: string
- name: rejected
dtype: string
splits:
- name: train
num_bytes: 1211199708
num_examples: 393999
- name: eval
num_bytes: 3029624
num_examples: 1000
download_size: 561354683
dataset_size: 1214229332
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: eval
path: data/eval-*
---

# Dataset Card for "MetaMath_DPO_FewShot"

GSM8K \citep{cobbe2021training} is a dataset of diverse grade school maths word problems, which has been commonly adopted as a measure of the math and reasoning skills of LLMs.
The [MetaMath](https://meta-math.github.io/) dataset is an extension of the training set of GSM8K using data augmentation.
It is partitioned into queries and responses, where the query is a question involving mathematical calculation or reasoning, and the response is a logical series of steps and calculations that culminate in a final answer.
To construct our paired-preference version of MetaMath, we take the queries as prompts x and the responses as the preferred completions y_w. We create y_l by modifying the response by randomly corrupting one of the results of an intermediate calculation so that it is incorrect; however, we leave the answer untouched.

An example is as follows:

x: "What is the total cost of purchasing equipment for all sixteen players on the football team, considering that each player requires a $25 jersey, a $15.20 pair of shorts, and a pair of socks priced at $6.80?"

y_w: "Each player requires a $25 jersey, a $15.20 pair of shorts, and a pair of socks priced at $6.80. So the total cost for each player is $25 + $15.20 + $6.80 = 47. Since there are sixteen players on the football team, the total cost for all of them is 16 * $47 = $752. \#\#\#\# 752 The answer is: 752"

y_l: "Each player requires a $25 jersey, a $15.20 pair of shorts, and a pair of socks priced at $6.80. So the total cost for each player is $25 + $15.20 + $6.80 = 52. Since there are sixteen players on the football team, the total cost for all of them is 16 * $47 = $752. \#\#\#\# 752 The answer is: 752"

Our motivation in building this dataset is to align models towards being precise in intermediate calculations.
This dataset has low edit distance -- the normalised edit distance is approximately 6.5%.
The dataset is meant to be used to fine-tune LLMs (which have already undergone SFT) using the DPOP loss function. We used this dataset to create the [Smaug series of models](https://github.com/abacusai/smaug).

The dataset contains 393,999 training examples and 1,000 evaluation examples.

See more details in the [datasheet](https://github.com/abacusai/smaug/blob/main/datasheet.md), and in our paper: https://arxiv.org/abs/2402.13228.

许可证：Apache-2.0
数据集信息：
特征项：
- 字段名：提示（prompt），数据类型：字符串
- 字段名：chosen，数据类型：字符串
- 字段名：rejected，数据类型：字符串
数据集划分：
- 划分名称：训练集（train），字节数：1211199708，样本数：393999
- 划分名称：评估集（eval），字节数：3029624，样本数：1000
下载大小：561354683字节
数据集总大小：1214229332字节
配置项：
- 配置名称：默认（default），数据文件：
- 训练集划分：路径为data/train-*
- 评估集划分：路径为data/eval-*


# 「MetaMath_DPO_FewShot」数据集卡片

GSM8Kcitep{cobbe2021training}是一个涵盖多样化小学阶段数学应用题的数据集，常被用作评估大语言模型（Large Language Model，LLM）的数学能力与推理能力的基准数据集。
[MetaMath](https://meta-math.github.io/)数据集是通过数据增强手段扩展GSM8K训练集得到的衍生数据集。该数据集被划分为查询与回复两部分：查询为包含数学计算或推理的问题，回复则为一套逻辑连贯的推导步骤与计算过程，最终得出最终答案。

为构建MetaMath的配对偏好版本，我们将查询作为提示（prompt），即输入x，将原回复作为优选完成项$y_w$。随后，我们通过随机篡改某一步中间计算结果使其出错，来生成劣质回复$y_l$，但保留最终答案不变。

示例如下：
> x: "若每位足球队员需配备一件25美元的球衣、一条15.20美元的球裤与一双6.80美元的球袜，那么为全队16名队员购置装备的总花费是多少？"
> $y_w$: "每位队员需配备一件25美元的球衣、一条15.20美元的球裤与一双6.80美元的球袜。因此单名队员的装备总花费为25美元+15.20美元+6.80美元=47美元。全队共有16名队员，因此总花费为16×47美元=752美元。####752 最终答案为：752"
> $y_l$: "每位队员需配备一件25美元的球衣、一条15.20美元的球裤与一双6.80美元的球袜。因此单名队员的装备总花费为25美元+15.20美元+6.80美元=52美元。全队共有16名队员，因此总花费为16×47美元=752美元。####752 最终答案为：752"

本数据集的构建初衷是使模型在中间计算环节具备更高的精准性。该数据集的编辑距离极低，归一化编辑距离约为6.5%。

本数据集旨在用于基于DPOP损失函数对已完成监督微调（Supervised Fine-Tuning，SFT）的大语言模型进行进一步对齐微调。我们曾使用该数据集构建[Smaug系列模型](https://github.com/abacusai/smaug)。

本数据集包含393999条训练样本与1000条评估样本。

更多细节可查阅[数据集说明文档](https://github.com/abacusai/smaug/blob/main/datasheet.md)与我们的论文：https://arxiv.org/abs/2402.13228。

原始信息汇总

数据集卡片 "MetaMath_DPO_FewShot"

数据集描述

GSM8K citep{cobbe2021training} 是一个包含多样化的中小学数学应用题的数据集，通常被用作衡量大型语言模型（LLMs）数学和推理能力的标准。MetaMath 数据集是GSM8K训练集的扩展，通过数据增强构建。

数据集结构

数据集分为查询和响应，其中查询是一个涉及数学计算或推理的问题，响应是一个逻辑推理和计算的步骤序列，最终得出答案。为了构建配对偏好版本的MetaMath，我们将查询作为提示x，将响应作为首选完成y_w。通过随机修改中间计算结果中的一个使其不正确来创建y_l，但保持答案不变。

示例

x: "What is the total cost of purchasing equipment for all sixteen players on the football team, considering that each player requires a $25 jersey, a $15.20 pair of shorts, and a pair of socks priced at $6.80?"

y_w: "Each player requires a $25 jersey, a $15.20 pair of shorts, and a pair of socks priced at $6.80. So the total cost for each player is $25 + $15.20 + $6.80 = 47. Since there are sixteen players on the football team, the total cost for all of them is 16 * $47 = $752. #### 752 The answer is: 752"

y_l: "Each player requires a $25 jersey, a $15.20 pair of shorts, and a pair of socks priced at $6.80. So the total cost for each player is $25 + $15.20 + $6.80 = 52. Since there are sixteen players on the football team, the total cost for all of them is 16 * $47 = $752. #### 752 The answer is: 752"

数据集目的

构建此数据集的目的是使模型在中间计算中更加精确。该数据集的编辑距离较低，标准化编辑距离约为6.5%。该数据集旨在用于使用DPOP损失函数对已经进行过SFT的LLMs进行微调。我们使用此数据集创建了Smaug系列模型。

数据集规模

数据集包含393,999个训练样本和1,000个评估样本。

更多详情

更多详细信息请参见datasheet和我们的论文：https://arxiv.org/abs/2402.13228。