DataArc SynData Toolkit
A modular, highly user-friendly synthetic data generation toolkit supporting multi-source, multi-language data synthesis.
Easily synthesize training data for LLMs with zero-code CLI and GUI !
:book: [ English | 中文 ]
:dart: Project Overview
DataArc SynData Toolkit is a synthetic data generation toolkit developed and open-sourced by DataArcTech and International Digital Economy Academy. It enables users to generate customized training data in one step through simple configuration files based on their requirements. The toolkit is described in our paper DataArc-SynData-Toolkit: A Unified Closed-Loop Framework for Multi-Path, Multimodal, and Multilingual Data Synthesis, published in the ACL 2026 System Demonstrations track (PDF).
:bulb: Key Features
- Extremely Simple Usage: Synthesize data with a single command and a configuration file. GUI is also provided for easy operations.
- Support for Multi-Source Synthetic Data:
- Local Synthesis: Support for generating data based on local corpora.
- Huggingface Integration: Automatically crawl and filter data from Huggingface.
- Model Distillation: Enable synthetic data generation through model distillation.
- Integrated Post-Training Module: End-to-end model training workflows powered by verl, supporting SFT and GRPO.
- Post-Training Model Evaluation: Evaluate trained models using DeepEval.
- Multilingual Support: Supports English and various low-resource languages.
- Multi-Provider Model Support: Works with local deployment, OpenAI APIs, and more.
- Highly Extensible: The entire synthetic data workflow is modular, allowing developers to flexibly customize them.
:movie_camera: Demo
Watch our 2-minute demo to experience how DataArc SynData Toolkit works in practice.
https://github.com/user-attachments/assets/4b4d5ae4-d274-4971-a3cb-e9f07e841374
We also provide a complete tutorial to help you get started.
:microscope: Performance
| Model | Medical | Finance | Law |
|---|---|---|---|
| Qwen-2.5-7B-Instruct | 42.34% | 52.91% | 19.80% |
| Trained with Synthetic Data | 64.57% | 73.93% | 42.80% |
A few lines of code deliver over 20% performance improvements.
:notebook: Changelog
[25/11/17] 🎉We open-sourced our synthetic data platform.
[25/11/27] We added parallel processing module to significantly accelerate the synthetic data generation pipeline.
[25/11/28] We added intermediate result saving, allowing users to resume from the last successful stage** instead of restarting the entire pipeline — a major token saver.
[25/12/25] 🔥Major upgrade:
- Frontend–Backend Separation: DataArc SynData Toolkit now adopts a fully frontend–backend separated architecture, featuring a FastAPI backend (REST APIs + SSE streaming for real-time progress) and a standalone React frontend for improved visualization, usability, and scalability.
- Post-Training Support via verl: Introduced an integrated post-training module powered by verl, enabling end-to-end model training workflows including SFT and GRPO on synthesized data.
- Multilingual Expansion: Added support for generating Arabic datasets, leveraging an Arabic translation model to produce fully localized synthetic data outputs.
[26/01/12] 🖼️ Image Modality Support Added:
- Image modality Local Task: Generate VQA (Visual Question Answering) data from local images or PDF-extracted figures using VLMs.
- Image modality Web Task: Automatically search and retrieve image-text datasets from HuggingFace Hub.
[26/01/26] 📊 Post-Training Model Evaluation:
- DeepEval Integration: Added model evaluation module powered by DeepEval.
- Three Evaluation Metrics:
- Answer Correctness: Compare model output against ground truth with customizable rubrics.
- Pairwise Preference: Compare post-trained model vs base model to measure improvement.
- Format Compliance: Evaluate adherence to output format instructions.
[26/09/09] 🎓 Our paper DataArc-SynData-Toolkit: A Unified Closed-Loop Framework for Multi-Path, Multimodal, and Multilingual Data Synthesis was published in the Proceedings of ACL 2026 (Volume 3: System Demonstrations). Read it on the ACL Anthology and see Citation.
[!TIP]
If you cannot use the latest feature, please pull the latest code.
:factory: DataArc SynData Toolkit Pipeline
DataArc SynData Toolkit is designed to synthesize data in a modular pipeline, allowing users to customize the strategies and implementation methods of each step. The main components include:
- Synthetic Data Generation: Generate data through methods such as local synthesis, Huggingface dataset retrieval, and model distillation.
- Developers can inherit BaseTaskConfig and BaseTaskExecutor to customize the generation task.
- Data Filtering and Rewriting: Filter and rewrite initially synthesized data according to the target model's requirements.
- Developers can inherit BaseRewriteConfig and BaseRewriter to customize the rewrite method for synthetic data (or no rewriting).
By decoupling modules, developers can achieve flexible customization of functional modules based on specific needs.
:jigsaw: Use Cases
We provide three different use cases that sythesize data through DataArc SynData Toolkit.
:file_folder: Project Structure
DataArc-SynData-Toolkit/
├── configs/ # YAML configuration examples
│ ├── sdg.yaml # SDG pipeline config
│ ├── sft.yaml # SFT training config
│ ├── grpo.yaml # GRPO training config
│ └── eval.yaml # Model evaluation config
│
├── sdgsystem/ # Core System
│ ├── app/ # FastAPI backend (REST + SSE)
│ ├── generation/ # Data generation
│ ├── documents/ # File parsing & retrieval
│ ├── huggingface/ # HF dataset integration
│ ├── distillation/ # Model distillation synthesis
│ ├── tasks/ # SDG execution tasks
│ ├── evaluation/ # Data quality evaluation
│ ├── deepeval/ # Post-training model evaluation (G-Eval)
│ ├── models/ # Unified LLM interface & postprocess
│ ├── trainer/ # Post-training (verl: SFT + GRPO)
│ ├── translation/ # Multilingual support
│ ├── webui/ # React frontend
│ ├── pipeline.py # Core SDG pipeline
│ └── cli.py # CLI entry
│
├── verl/ # Integrated verl framework
├── docs/ # Documentation
├── pyproject.toml
├── LICENSE # Apache License 2.0
└── README.md
:rocket: Quick Start
1. Install DataArc SynData Toolkit
# 1. Clone the repository
git clone https://github.com/DataArcTech/DataArc-SynData-Toolkit.git
cd DataArc-SynData-Toolkit
# 2. Install uv if not already installed
pip install uv
# 3. Install dependencies
uv sync
For hardware requirements and dependencies detail, please refer to dependency and installation guide.
2. Configuration
Please refer to the example configuration file and modify the configuration based on your requirements.
3. Synthesize Data
Run through CLI:
Create a .env file and specified the following fields.
API_KEY=sk-xxx # your api key
BASE_URL=https://api.openai.com/v1 # Optional: your base url
And run following command.
uv run sdg generate configs/sdg.yaml # or change to your .yaml file
:twisted_rightwards_arrows: Training with Synthesized Data
DataArc SynData Toolkit integrates an end-to-end model training module powered by verl, enabling you to train models directly on your synthesized data. We support two training methods: SFT (Supervised Fine-Tuning) and GRPO (Group Relative Policy Optimization).
Quick Start with CLI
1. Prepare Your Configuration
Create a training configuration file based on the SFT Configuration Example or GRPO Configuration Example.
2. Run Training
# SFT training
uv run sdg train configs/sft.yaml
# GRPO training
uv run sdg train configs/grpo.yaml
For detailed configuration options, refer to the example YAML files.
:bar_chart: Evaluating Post-Trained Models
DataArc SynData Toolkit provides a model evaluation module powered by DeepEval, enabling you to assess your post-trained models using LLM-as-a-Judge (G-Eval). We support three metrics: Answer Correctness, Format Compliance and Pairwise Preference.
Quick Start with CLI
1. Prepare Your Configuration
Create an evaluation configuration file based on the Evaluation Configuration Example.
Add your API keys to .env file.
OPENAI_API_KEY=sk-xxx # your openai api key
OPENAI_BASE_URL=https://api.openai.com/v1 # Optional: your openai base url
CONFIDENT_API_KEY=confident_us_xxx # your confident ai api key (access deepeval, can be created for free after signup)
2. Run Evaluation
uv run sdg eval configs/eval.yaml
Results can be visulized on confident ai webpage and will be saved to the configured output directory.
:desktop_computer: Run with GUI
Start FastAPI server with following command.
uv run fastapi dev sdgsystem/app/main.py
Open another terminal and build frontend with following command.
cd sdgsystem/webui
# Install dependencies
pnpm install
# Start development server
pnpm dev
If you have any doubt about regrading our Web UI, check our Web UI document.
:date: Schedule for the Next Release
- Encrypted Synthetic Data Generation: Run generation with sensitive data encrypted.
:handshake: Contributing
We welcome contributions!
:page_facing_up: Citation
If you find DataArc SynData Toolkit useful in your research or products, please cite our ACL 2026 System Demonstrations paper:
@inproceedings{shi-etal-2026-dataarc,
title = "{D}ata{A}rc-{S}yn{D}ata-Toolkit: A Unified Closed-Loop Framework for Multi-Path, Multimodal, and Multilingual Data Synthesis",
author = "Shi, Zhichao and
Yang, Cehao and
Zhou, Hao and
Wu, Xiaojun and
Li, Huajie and
Jiang, Xuhui and
Xu, Chengjin and
Wang, Yuanzhuo and
Guo, Jian",
editor = "Durrett, Greg and
Jian, Ping",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 3: System Demonstrations)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-demo.31/",
doi = "10.18653/v1/2026.acl-demo.31",
pages = "318--326",
ISBN = "979-8-89176-392-0"
}
:scroll: License
DataArc SynData Toolkit is released under the Apache License 2.0. The verl training framework bundled in the verl/ directory is redistributed under its original Apache License 2.0 (Copyright Bytedance Ltd. and/or its affiliates). See NOTICE and THIRD_PARTY_NOTICES.md for attribution details.