Easy Dataset
A powerful tool for creating datasets for LLM fine-tuning 、RAG and Eval
Last verified:
What is Easy Dataset?
Easy Dataset is a powerful open-source tool for creating fine-tuning datasets for Large Language Models (LLMs). It provides a complete closed-loop workflow from literature parsing to dataset construction, annotation, export, and evaluation. The tool is designed to solve the challenging dataset preparation stage of LLM fine-tuning by offering intelligent document processing, automated question generation, and answer construction with Chain-of-Thought support.
Key features include intelligent document processing supporting PDF, Markdown, DOCX, and TXT formats with multiple parsing options including MinerU API for complex PDFs. It uses Section-Aware Recursive Chunking algorithm for semantic-level text segmentation, AI-assisted batch question generation with interruption recovery, and intelligent answer construction that supports reasoning models like DeepSeek-R1 for COT generation. The tool includes a domain tag system with two-level domain trees, multi-format export (Alpaca, ShareGPT, multilingual-thinking in JSON/JSONL), built-in evaluation system with model testing and blind-test features, and a dataset marketplace aggregating HuggingFace and Kaggle data sources.
Easy Dataset is ideal for AI developers, data scientists, and researchers who need to build high-quality domain-specific datasets for LLM fine-tuning, RAG systems, and model evaluation. It suits both technical and non-technical users with its intuitive visual interface, supporting users who want to convert domain knowledge into structured datasets without manual data preparation.
Easy Dataset pricing
Pricing model: Freemium
Easy Dataset is an open-source tool under AGPL 3.0 license, available for free. It can be run locally via npm (npm install, npm run build, npm run start), using official Docker image (ghcr.io/conardLi/easy-dataset), or by building from local Dockerfile. The application runs on http://localhost:1717. Users need to configure their own LLM API keys (OpenAI, DeepSeek, Ollama, etc.) which are billed separately by the respective providers. MinerU API Key is required for advanced PDF parsing and can be applied for free at https://mineru.net/apiManage/token with 14-day validity requiring reconfiguration.
Easy Dataset pros
- Open-source with AGPL 3.0 license
- Supports multiple document formats: PDF, Markdown, DOCX, TXT
- MinerU API parsing for complex PDFs with formulas and charts
- Section-Aware Recursive Chunking algorithm for semantic segmentation
- Batch question generation with interruption recovery
- Supports Chain-of-Thought (COT) generation for reasoning models
- Compatible with all OpenAI-format LLM APIs including OpenAI, DeepSeek
- Supports local models via Ollama
- Built-in Playground for model testing and multi-model comparison
- Two-level domain tag system with AI automatic generation
- Multiple export formats: Alpaca, ShareGPT, multilingual-thinking
- Exports to JSON and JSONL file types
- Built-in evaluation system with model testing and blind-test features
- Dataset marketplace with HuggingFace and Kaggle integration
- Docker deployment option for quick setup
- Visual interface for editing questions, answers, and datasets
- Custom system prompts support
- Configurable concurrency settings for batch tasks
- Manual editing and AI optimization of Q&A pairs
- 11.3k GitHub stars showing strong community adoption
Easy Dataset cons
- Built-in PDF parsing less effective than MinerU for complex files
- MinerU API key validity only 14 days, requires reconfiguration
- Reasoning models slow down text segmentation processing
- Batch tasks interrupted by closing or refreshing page
- Requires model selection before text segmentation or processing fails
- PDF processing time longer with MinerU API or custom vision model
- Custom vision model parsing requires additional model configuration
- Maximum concurrency limited by model provider restrictions
- JavaScript-heavy (99.5%) may limit customization for non-JS developers
- No explicit free tier pricing mentioned, self-hosted only
Frequently asked questions about Easy Dataset
What is Easy Dataset and what problem does it solve?
Easy Dataset is a powerful large model dataset creation tool that provides a systematic solution implementing a complete closed-loop from literature parsing to dataset construction, annotation, export, and evaluation. It solves common dataset building problems including lack of knowledge on how to proceed, poor AI performance on large files, context limitations causing truncation, duplicate questions in batch generation, managing compiled datasets for annotation, building domain tags, constructing Chain-of-Thought for reasoning models, and converting between dataset formats.
What document formats does Easy Dataset support?
Easy Dataset supports processing literature in four formats: Markdown, PDF, DOCX, and TXT. Markdown files are recommended as models understand them best with good structural organization. For PDFs specifically, there are four parsing methods: Basic Parsing for simple files, MinerU API Parsing for complex PDFs with formulas and charts, MinerU Online Platform Parsing, and Custom Vision Model Parsing for complex PDF recognition.
Which LLM models and APIs are supported?
Easy Dataset supports all OpenAI format APIs including OpenAI, DeepSeek, and various third-party model providers. It also supports local models through Ollama. The Model Configuration Center includes a built-in Playground for model testing and supports multi-model comparison. Users should select normal question-answering models like Doupai or Qianwen for text segmentation, not reasoning models.
How does batch question generation work?
Batch question generation is based on text block semantics and dynamically generates questions according to character density (configurable). It supports batch creation and interruption recovery. The process runs in the background, unlike foreground execution which interrupts if you switch tabs. You can control the maximum number of concurrent tasks in Task Settings - Question Generation Settings, where larger concurrency means faster generation but must respect the model provider's concurrency limits.
How is Chain-of-Thought (COT) generated?
When a reasoning model like DeepSeek-R1 is selected in the upper right corner during dataset generation, the chain of thought in the model's reasoning process is preserved. Easy Dataset generates answers based on the question plus the text block corresponding to the question plus domain tags together, ensuring answer relevance to the literature itself. The COT information is included in multi-format exports.
What export formats are available?
Easy Dataset supports multiple export formats including Alpaca, ShareGPT standard formats, and multilingual-thinking. File types include JSON and JSONL. The export supports custom field mapping, including domain tags and COT information. Users can also add custom system prompts during export to guide model responses.
How do I deploy Easy Dataset locally?
There are three deployment methods: 1) Install with NPM - clone repository, run 'npm install', then 'npm run build' and 'npm run start', visit http://localhost:1717. 2) Using Official Docker Image - clone repository, modify docker-compose.yml for paths, run 'docker-compose up -d', visit http://localhost:1717. 3) Building with Local Dockerfile - clone repository, 'docker build -t easy-dataset .', then run container with port 1717 and volume mounts for local-db and prisma.
What is the domain tag system?
The domain tag system uses AI to automatically generate two-level domain trees (such as 'Sports-Football'), supports manual correction, and binds precise tags to each Q&A pair, reducing duplication rate. The system has global understanding capabilities and helps organize datasets according to domain trees, making it easier to build domain-specific tags for specific domain requirements.
How does quality verification work?
Easy Dataset provides a quality verification mechanism including batch deletion, manual editing, and AI optimization of questions and answers. AI optimization allows automatic polishing with input instructions to ensure data usability. Users can filter questions with generated answers and questions without generated answers, view dataset details, and edit questions, answers, and datasets at any stage of the process.
What is the Dataset Marketplace feature?
The Dataset Marketplace aggregates multiple platform data sources such as HuggingFace and Kaggle, supports one-click keyword search, and solves the initial problem of 'where to get data.' This feature helps users who need additional information or want to start with existing datasets rather than building from scratch, providing a convenient way to discover and access pre-existing datasets for their fine-tuning needs.