Data-to-Paper

What is Data-to-Paper?

Data-to-Paper is an open-source automation framework that guides interacting language-model and rule-based agents through an end-to-end scientific research process. Starting with raw data, it explores the dataset, searches literature, proposes and tests hypotheses, writes and debugs analysis code, interprets results, and produces a complete research paper.

The framework emphasizes transparent, backward-traceable results. Numeric claims in generated manuscripts can be traced back to the code that produced them. It can run autonomously or through a Copilot app that lets researchers set goals, inspect and guide work, review outputs, rewind steps, record and replay runs, and monitor API costs.

How to Use Data-to-Paper

1. Install the Python package and its documented dependencies.
2. Run the data-to-paper command and provide a supported dataset or research case.
3. Choose autonomous operation or use Copilot controls to set goals, inspect progress, review work, and rewind steps.
4. Examine the generated analysis and trace manuscript values back to their source code.
5. Have a domain expert verify the paper's rigor, accuracy, ethics, and conclusions before use.

Core Features

• End-to-end research automation: Covers exploration, literature search, hypothesis generation, analysis, interpretation, and paper writing.
• Backward-traceable manuscripts: Connects numeric statements to the analysis code that generated them.
• Autopilot and Copilot modes: Supports autonomous runs or human-guided research.
• Review and rewind controls: Lets users inspect, guide, review, replay, and return to earlier steps.
• Coding guardrails: Overrides standard statistical packages with safeguards intended to reduce common model-generated coding errors.
• API cost tracking: Helps users monitor model usage and related costs.

Use Cases

• Exploratory data research: Moves from an unfamiliar dataset to candidate questions and analyses.
• Hypothesis testing: Raises and tests statistical hypotheses for relatively simple research goals.
• Reproducible paper generation: Produces manuscripts whose results can be traced to code.
• AI research evaluation: Helps study the capabilities and limitations of autonomous scientific agents.

Pricing

Data-to-Paper is distributed under the MIT license. The repository does not charge a software subscription, but users are responsible for costs from external language-model APIs.

Frequently Asked Questions

Is human review still required?

Yes. The maintainers state that the process is not error-proof and that domain-expert oversight is necessary.

Can it run without continuous guidance?

Yes. It supports autonomous operation as well as a human-guided Copilot mode.

What kind of research is it designed for?

The repository says it currently targets relatively simple research goals and datasets involving statistical hypotheses.