data-to-paper
An open-source AI research framework that runs analysis code locally, uses external LLM APIs, and links paper results to the code that produced them.
data-to-paper is an open-source Python framework for researchers who want AI agents to work through a dataset and draft a scientific paper whose numerical claims they can check. Its defining feature is traceability: readers can click a value in the manuscript and follow it back to the specific analysis code that produced it.
The framework covers data exploration, literature searches, hypothesis development, analysis and interpretation, then writes the paper. It combines LLM agents with rule-based agents to guide that process. Generated analysis code runs on your machine, while language model calls use external APIs, including GPT-4. It isn't fully offline.
Researchers can let the agents propose and test hypotheses autonomously or use the Copilot App to set goals, inspect their work and guide decisions. The app supports human feedback and on-demand AI reviews. You can also return to an earlier research step, record and replay runs, and track API costs.
The intended scope is relatively simple datasets and research questions centered on statistical hypotheses. Examples cover health indicators, social network data and clinical treatment questions. Guardrails around statistical packages aim to reduce common mistakes in AI-written analysis code, but the results still need a domain expert's review.
The code uses the MIT license. Generated manuscripts carry a watermark identifying them as AI-created.
Logotyper
