An open taxonomy, dataset, and set of benchmarks to help AI understand why something is funny — across cultures, formats, and contexts. A project incubated by Midtown Show.
Curated clips, transcripts, and annotations labeled by humor type, setup→punch structure, and audience reaction timing.
A research-backed schema unifying comedic devices (irony, misdirection, call-back), formats, tones, and social context.
Standard tasks for detection, generation, and alignment — with reproducible protocols and baselines.
Researchers, comedians, and curators co-designing annotation guidelines and culturally aware test sets.
Why now
Humor is context, culture, and timing. We’re building shared infrastructure — a transparent taxonomy, open datasets, and strong benchmarks — so models can learn structure, not just style. Our north star: safer, more respectful, culturally aware comedic understanding that supports human creators.
Tracks
Consensus-driven schema for humor categories and devices with clear annotation rules and edge-case guidance.
Multilingual, multi-format dataset with clips, transcripts, and crowd+expert labels. Opt-in ethics and creator credit.
Detection, classification, and generation tasks; leaderboards and baselines for fair comparison.
Open-source scripts for segmentation, laugh-detection, and timing alignment; lightweight SDK for experiments.
Get involved
Sign up for early contributor updates, annotation pilots, and kickoff calls. We’ll also announce the initial taxonomy draft and submission guidelines here.