A benchmark for Artificial Pub Intelligence

Take a seat. Identify the pub.

One interior photo. One hint: country: GB. Can a vision model name the exact pub?

protocol v0 · 0 / 200 pubs · results: none yet

A pub garden bench with two answer slips, a BenchPub plaque, and a beer mat under one leg

Noticeboard

  • Collection open. Submit an interior photo of a UK pub.
  • Pilot in progress. 50 pubs × 5 models × 2 tracks × 3 runs. 1,500 attempts to recognise some carpet.
  • Quiz and results board. After the pilot. Not before.

The task

"This photograph was taken inside a UK pub. Identify the exact establishment, or abstain."

No town. No shortlist. No reverse-image search. The answer must resolve to one venue: "The Red Lion" is not an answer, it is a genre. Scored as exact-venue accuracy; abstentions and failures stay in the denominator.

Protocol, tracks, scoring →