Config-driven multimodal data processing pipeline for MLLMs

---

简体中文 | Documentation

---

Introduction

DataStudio is a data processing toolkit for Multimodal Large Language Models (MLLMs). It supports both rule-based and MLLM-powered operators for filtering and rewriting data. DataStudio drives the HoneyPipe data cleaning pipeline that produced Honey-Data-15M — a high-quality dataset of 15 million QA pairs.

Key Features

• Config-Driven: Define complete processing pipelines via Python config files with config inheritance (_base_)
• Dual-Engine Operators: 16 built-in rule operators (7 filters + 9 rewriters), plus MLLM-powered operators for intelligent filtering and rewriting
• High Performance: Multi-process async concurrent API requests (8192+ via MPOpenAIAPI), LMDB sharded image caching, fork + COW memory sharing
• Checkpoint Resume: Automatic progress saving with checkpoint-based recovery after interruption

Auxiliary Tools

• LLMRouter: A high-performance reverse proxy router built in Go, OpenAI API-compatible, supporting multiple LLM backends with weighted load balancing, sliding-window RPM throttling, async health checks with auto failover, and config hot-reload. See its README for details.
• DataVis: A multimodal data visualization and analysis platform (React + FastAPI) for browsing, filtering, and analyzing processed datasets.

---

Installation

git clone https://github.com/Open-Bee/DataStudio.git
cd DataStudio
pip install -e .

Requirements: Python >= 3.10.

---

Quick Start

1. Prepare a Dataset YAML

# dataset.yaml
datasets:
  - json_path: /path/to/data.jsonl
    source: my_dataset

2. Write a Config File

# configs/my_pipeline.py
_base_ = ["@/_base_/dataset.py"]

work_dir = "./work_dirs/my_experiment"
logger = dict(type="Logger", log_file="logs/process.log")
dataset_yaml = "/path/to/dataset.yaml"

dataloader = dict(dataset=dataset_yaml, batch_size=1000, use_image=False)
datasaver = dict(dataset=dataset_yaml, output_dir="./output", save_yaml_name="processed")

pipeline = dict(
    type="Pipeline",
    operations={
        "rule_filters": dict(
            cfg=dict(type="SubPipeline", operators=[
                dict(type="ConvLengthFilter", min_length=1, max_length=20),
                dict(type="ImageSizeFilter", min_size=14),
            ]),
            priority=0,
        ),
        "rewriters": dict(
            cfg=dict(type="SubPipeline", operators=[
                dict(type="RemoveThinkRewriter"),
            ]),
            priority=1,
        ),
    }
)

3. Run

python run.py -c configs/my_pipeline.py

For a complete tutorial covering data formats, MLLM operators, custom operators, performance tuning, and more, see the Quick Start Guide.

---

Project Structure

DataStudio/
├── run.py                  # Main entry point (supports -c/--config, --cache-images)
├── datastudio/
│   ├── operators/          # Operators
│   │   ├── core/           # Base: Operator, Filter, Rewriter, DataItem, OperatorResult
│   │   ├── filters/        # 7 rule-based filters
│   │   ├── rewriters/      # 9 rule-based rewriters
│   │   └── mllm/           # MLLM operators: MLLMFilter, MLLMRewriter, SelectiveMLLMRewriter
│   ├── models/             # Model backends: OpenAIAPI, MPOpenAIAPI
│   ├── pipelines/          # Pipeline, SubPipeline
│   ├── datasets/           # Data loading/saving, format conversion, YAML generation
│   └── utils/              # Registry, config, logging, LMDB cache, checkpoint
├── configs/                # Configuration files
│   ├── _base_/             # Base configs (models, datasets, filters, rewriters)
│   └── examples/           # Ready-to-run example configurations with demo data
├── prompts/                # MLLM prompt templates
├── tools/                  # Auxiliary tools
│   ├── LLMRouter/          # OpenAI API-compatible reverse proxy with load balancing
│   └── DataVis/            # Multimodal data visualization and analysis platform
├── docs/                   # Documentation (Sphinx + Markdown guides)
└── tests/                  # Unit tests

---

About the Bee Project

DataStudio is the data cleaning engine of the Bee project. Through the HoneyPipe pipeline, it produced the Honey-Data-15M dataset.

---

Citation

@article{zhang2025bee,
  title={Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs},
  author={Zhang, Yi and Ni, Bolin and Chen, Xin-Sheng and Zhang, Heng-Rui and Rao, Yongming and Peng, Houwen and Lu, Qinglin and Hu, Han and Guo, Meng-Hao and Hu, Shi-Min},
  journal={arXiv preprint arXiv:2510.13795},
  year={2025}
}

Contributing

If you encounter any bugs or have feature requests, feel free to open an Issue or submit a Pull Request. Contributions are welcome!

License

Apache License 2.0