Easy Dataset
An application for building datasets to fine-tune, retrieval-augment and evaluate large language models, with document parsing, text splitting, question generation and data cleaning built in.
Easy Dataset is an application for building datasets used to fine-tune, retrieve-augment and evaluate large language models. It parses documents in formats such as PDF, Markdown, DOCX, TXT and EPUB, splits them into segments using several selectable algorithms, and automatically extracts questions from each segment with customizable templates. It then builds a global domain label tree from the document structure, uses an LLM API to generate answers and chain-of-thought reasoning, and cleans the resulting text to remove noise.
It can produce several kinds of dataset: single-turn question-answer pairs, multi-turn dialogues with customizable roles and scenarios, image question-answer pairs built from images supplied as a directory, PDF or archive, and distilled label data.
A newer version adds evaluation features: it can convert domain documents into test sets and run multi-dimensional evaluation, including a human blind-test system for comparing model or fine-tuning results.