The dataset file featherweight_v2.jsonl hasn't been uploaded to this site yet. The online test will activate as soon as it's available.
In the meantime, the documentation describes the methodology and human baseline.
Couldn't load the test.
Featherweight v2 · Take the test
How would you do on a small-model benchmark?
Same 300 questions our models see. Same scoring as the paper runs that produced the ~86.8% human baseline. Your answers are stored only in your browser; no scores are sent anywhere. Read the full methodology for details.
300
Questions
10
Categories
~86.8%
Human baseline
86.7%
Epos 1
Session in progress
You have an unfinished test.
Pick a mode
No timer, no penalty for wrong answers. Guess if you don't know — that's what the paper participants were instructed to do. Your final score is the count of correct letters out of total questions.