Minecraft_QA-pairs_Instruction_Dataset

Overview

This repository contains a dataset designed for training large language models (LLMs) to handle a variety of questions related to Minecraft. Our dataset includes over 390,000 instruction entries generated using an LLM-assisted method. The dataset is designed to provide high-quality, diverse Q&A pairs, which are essential for effective model training.

Dataset Generation

Data Collection and Cleaning

We utilized two primary sources of information:

Minecraft Fandom Wiki (https://minecraft.fandom.com/wiki/Minecraft_Wiki)

Minecraft Wiki (https://minecraft.wiki/)

For both sources, we performed extensive data cleaning to remove irrelevant sections, such as game version information, achievements, advancements, and history. The remaining content was extracted and stored in text (TXT) and markdown (MD) files, ensuring that only useful text and table content were retained.

Q&A Pair Generation

Based on the cleaned data, we employed GPT-3.5-Turbo~\citep{chatgpt} with customized prompts to automatically generate diverse Q&A pairs. These pairs are categorized into four types:

Short Answer Q&A: Questions that can be directly answered with a word or phrase.

Long Answer Q&A: Questions that require one or more complete sentences for a comprehensive response.

Boolean Answer Q&A: Questions that can be answered with a simple True or False.

Normal Answer Q&A: General Q&A pairs with no specific form restrictions, ensuring the dataset's comprehensiveness.