Easy Dataset is a tool for creating fine-tuning datasets for large language models that converts domain-specific documents into structured data for model training, retrieval-augmented generation, and evaluation.
The tool addresses the challenge of preparing high-quality training data by automating document processing and dataset generation. It ingests documents in multiple formats including PDF, Markdown, DOCX, TXT, and EPUB, then applies intelligent text segmentation using various algorithms such as Markdown structure recognition, recursive separators, and code-aware chunking. The system can automatically generate questions from text segments using templates and batch processing, build domain label trees based on document structure, and generate answers with Chain of Thought reasoning through LLM APIs. Built-in data cleaning removes noise to improve dataset quality.
The tool supports multiple dataset output formats including single-turn question-answer pairs, multi-turn dialogue datasets with customizable roles and scenarios, and image-based QA data. It also includes a data distillation mode that generates label trees and questions directly from domain topics without requiring document uploads. For teams evaluating adoption, Easy Dataset suits projects where domain-specific fine-tuning data needs to be generated at scale from existing documentation. It is particularly valuable when you have large collections of unstructured documents and need to convert them into training data quickly, or when you need to create evaluation datasets and run multi-dimensional assessment tasks on fine-tuned models.
The project shows active development with regular feature additions including a recently launched evaluation system that enables automated multi-dimensional evaluation tasks and human blind testing capabilities for assessing model performance across vertical domains and RAG recall rates. The codebase is maintained in JavaScript and backed by comprehensive documentation with tutorials demonstrating core workflows.
