# peekaboolean

A small vision-language model, fine-tuned with LoRA on a 500-million-parameter backbone, that answers typed multiple-choice, ordinal-score and yes/no questions about an image with calibrated probabilities, in about 400 milliseconds on a laptop.

peekaboolean is a small vision-language model that answers typed questions about an image rather than generating free text, so it only ever returns answers to the specific choice, score or yes/no questions the caller supplied, with calibrated probabilities attached. It is built on the SmolVLM-500M-Instruct backbone with a LoRA adapter trained on the language model while the vision tower stays frozen, distilled from a larger local teacher model plus public visual-question-answering data.

A request sends one image, a short context, and any number of named questions, and each question is scored independently so answers do not depend on the order the questions were asked in. On an Apple M1 Pro it answers a six-question request in about 400 milliseconds, and about 60 milliseconds on a desktop GPU; the released weights are distributed under a non-commercial license, and the project documents what did and did not work while building it.

Website: https://github.com/bykof/peekaboolean