The same 1.5B distilled student as the
gallery, running entirely in your browser on your GPU
(WebLLM, WebGPU, 4-bit). Pick demographics,
generate a fresh persona, reasoning trace, and 24-hour activity chain —
no server, nothing you do here leaves your machine.
What to expect. The first load downloads ~1 GB of model weights
(cached by your browser afterwards). Outputs stream token by token and are shown unedited; a
small model occasionally produces a schema-invalid day, which we flag instead of hiding.
Requires WebGPU — Chrome or Edge 113+ on desktop (Safari 18+ and Firefox Nightly may work).
WebGPU not available in this browser.
You can still browse the precomputed gallery, which needs no GPU.
Schema-invalid generation.
This sample did not parse into a valid home-anchored day — that happens occasionally with a
1.5B model. The raw output is below; hit Generate to try again.
raw model output (streams live)
How this differs from the gallery
The gallery shows precomputed generations from our lab GPU. This page runs the identical
student model (Qwen2.5-1.5B + merged LoRA, one per city) in your browser, quantized to 4-bit
(q4f16_1) and decoded at the same temperature 0.8 / top-p 0.92 used throughout the paper. The
demographic vocabulary matches the training surveys, and the prompt sent to the model is exactly
the training-time format — what you see is the deployment artifact, not a hosted API.
Model weights are published at
huggingface.co/UMN-Choi-Lab (CC-BY-NC-4.0,
research use). Survey microdata is never published; profiles you compose here are synthetic by
construction.