Skip to Main Content
trunchbull

SkateBench

Trunchbull port

A playful, focused test of whether AI understands skateboard tricks and can follow a precise product recommendation task.

3 case previewschatMITversion a4d54c2

Run on your modelsTry it now, no account needed

Runs in the standard chat lane with an accepted-answer-bank grader.

What it measures

SkateBench asks a deceptively human question: does the model really understand the language of skateboarding, or is it guessing from familiar words? The Trunchbull CI port keeps the benchmark small, clear, and easy to inspect, including the separate UploadThing recommendation task.

SkateBench tests small but precise pieces of domain knowledge. Cases ask a model to identify a trick from a movement description or choose the expected upload recommendation.

The Trunchbull port preserves all eight upstream cases and grades responses against accepted answer banks, with forbidden phrases checked before accepted answers.

trunchbull port

Trunchbull maintains a complete eight-case port with accepted and forbidden answer rules.

Cases

3 in this release
idcaseruntime
skatebench/tre-flip
Tre flipIdentify a trick from its combined rotation and flip.
chat
skatebench/hardflip
HardflipRecognize a frontside shove-it and kickflip combination.
chat
skatebench/upload
Safe video uploadRecommend the expected upload product without extra claims.
chat

Tools

none required

No model tools. Cases are answered from the prompt and checked by the grader.

SkateBench, explained

What does SkateBench test?

SkateBench tests specialized skateboard terminology, trick identification, and a concise domain-specific recommendation task.

How many SkateBench cases are included?

The Trunchbull CI port preserves all eight upstream cases and evaluates them with accepted-answer and forbidden-answer rules.