Prompts to Proxies · v0.1.0

P2P documentation

P2P reconstructs how a human population answers a survey by building a pool of LLM proxy respondents and weighting them so their combined answers match the population's published answer shares. It needs only aggregate survey results, not respondent-level demographic data or model fine-tuning.

Get started Experiment Studio GitHub

One YAML configuration system and one command-line interface cover persona generation, survey elicitation, constrained regression, evaluation, sweeps, clustering and aggregation. The Experiment Studio (p2p ui) puts the same workflow in a local browser interface.

The quickstart, the fixture workflows and the test suite run offline. Provider backends (OpenAI, Gemini, DeepSeek) are optional, and Ollama runs the real pipeline on your own machine.

Where to start

01QuickstartInstall P2P and run the whole pipeline offline on a fictional survey. 02ConceptsThe two stages, the basis and target, and what every run records. 03Experiment StudioBuild, run and read experiments in a local browser interface. 04Command lineEvery p2p command, its options and examples. 05ConfigurationThe YAML manifest: data, Stage 1, Stage 2 and evaluation. 06API referenceModules, classes and functions, generated from the code.

The repository README.md gives the installation, the data sources, the relation to the NeurIPS 2026 paper Prompts to Proxies and how to cite P2P.