中文 | English
Data-Centric Autonomous AI Research
Figure 1. Overview of the DataMaster autonomous data-engineering framework.
DataMaster focuses on the data side of machine learning problem solving. Given a fixed modeling algorithm or starter solution, it searches for better data pipelines, external data sources, feature transformations, validation signals, and reusable data artifacts. It targets both MLE-Bench (competition-style ML tasks) and PostTrainBench (post-training enhancement: math, reasoning, domain-specific fine-tuning).
The framework organizes data-engineering decisions with a DataTree. Red nodes explore external data or transformations, Black nodes exploit and refine candidates, the Data Pool stores reusable datasets, and Global Memory retains outcomes across search rounds. DataMaster is built on top of EvoMaster.
DataMaster is designed to support both MLE-Bench and PostTrainBench workflows. The current open-source release includes the MLE-Bench workflow:
| Component | Path |
|---|---|
| DataTree core workflow | playground/data_master |
| Baseline MLE-Bench workflow | playground/ml_master |
| Dataset search & submission tools | playground/search_dataset_tools |
| Benchmark integration | mle-bench/ |
| Task-specific configs (75 tasks) | configs/data_master/ |
PostTrainBench is on the DataMaster roadmap. Code will be released in a future update.
|
|
Figure 2. Walkthrough of the DataTree search process: Red nodes explore, Black nodes exploit.
DataMaster/
├── configs/
│ ├── ml_master/ # Baseline agent configs
│ └── data_master/ # DataTree configs + 75 task configs
├── docs/ # English and Chinese documentation
├── evomaster/ # EvoMaster core components
├── initial_code/ # Starter-code templates
├── mle-bench/ # Benchmark integration and tooling
├── playground/
│ ├── ml_master/ # Baseline MLE-Bench workflow
│ ├── data_master/ # Main DataTree workflow
│ └── search_dataset_tools/ # Dataset search & submission tools
├── scripts/ # Utility and visualization scripts
├── run.py # Main CLI entry point
├── pyproject.toml
├── requirements.txt
└── LICENSE
git clone https://github.com/zhifan-zhou/DataMaster.git
cd DataMaster
python -m venv .venv
source .venv/bin/activate
pip install -e .Note: Python >= 3.10, < 3.13 required.
Configs live under configs/ml_master/ and configs/data_master/. Task-specific YAML configs are under configs/data_master/yaml_configs/, with MCP tool configs under configs/data_master/json_configs/.
Credentials are not stored in this repository. Provide keys via environment variables:
export DATA_ROOT=/path/to/mle-bench-lite
export LLM_MODEL=your-model-name
export LLM_API_KEY=your-api-key
export LLM_BASE_URL=https://your-llm-endpoint/v1
export SERPER_API_KEY=optional-serper-key # web search
export HF_TOKEN=optional-huggingface-token # dataset searchexport DATA_ROOT=/path/to/mle-bench-lite
export LLM_MODEL=your-model-name
export LLM_API_KEY=your-api-key
export LLM_BASE_URL=https://your-llm-endpoint/v1
python run.py \
--agent data_master \
--config configs/data_master/yaml_configs/detecting-insults-in-social-commentary/config_detecting-insults-in-social-commentary.yaml \
--task "$DATA_ROOT/detecting-insults-in-social-commentary/prepared/public/description.md"With a starter solution:
python run.py \
--agent data_master \
--config configs/data_master/yaml_configs/detecting-insults-in-social-commentary/config_detecting-insults-in-social-commentary.yaml \
--task "$DATA_ROOT/detecting-insults-in-social-commentary/prepared/public/description.md" \
--initial-code initial_code/data_loader_format/detecting-insults-in-social-commentary/full_code.py| Script | Purpose |
|---|---|
run.py |
Main CLI entry point |
scripts/auto_config_exp.py |
Generate task-specific configs |
scripts/build_full_initial_codes.py |
Assemble starter-code manifests |
scripts/prefetch_models.py |
Local model prefetch helper |
scripts/check_port_conflicts.py |
Diagnose grading-server port conflicts |
scripts/vis_node_by_tree_with_grade.py |
Interactive DataTree visualization |
This repository does not include MLE-Bench datasets, Kaggle data, model checkpoints, or generated artifacts. Prepare MLE-Bench / MLE-Bench Lite separately and point DATA_ROOT to the local benchmark directory. The vendored mle-bench/ directory provides integration code and reference tooling.
| Item | Status |
|---|---|
| MLE-Bench / MLE-Bench Lite workflow | ✅ Released |
| PostTrainBench workflow | 🔜 Coming soon |
| Additional documentation and examples | 🚧 In progress |
| Reproducibility scripts and benchmarks | 📋 Planned |
No credentials are intentionally included. Do not commit API keys, tokens, webhooks, SSH keys, .env files, benchmark data, model checkpoints, run logs, or private service configuration.
If you find DataMaster useful in your research, please cite:
@article{du2026datamaster,
title = {DataMaster: Towards Autonomous Data Engineering for Machine Learning},
author = {Yaxin Du and Xiyuan Yang and Zhifan Zhou and Wanxu Liu and
Zixing Lei and Zimeng Chen and Fenyi Liu and Haotian Wu and
Yuzhu Cai and Zexi Liu and Xinyu Zhu and WenHao Wang and
Linfeng Zhang and Chen Qian and Siheng Chen},
journal = {arXiv preprint arXiv:2605.10906},
year = {2026}
}DataMaster builds on EvoMaster and reuses its core agent and runtime abstractions. We thank the EvoMaster project for the upstream framework.
This repository is released under the Apache License 2.0.