What
JevNames is an experiment to put the model Jev to work as your partner in a Codenames-style game. Since Jev returns a probability over a set of options, it seemed perfect to put the model through its paces interpreting a human's clues.
How
9 words highlighted in green are your key words . You want Jev to guess all 9 of these to win.
1 word in black is the bomb . If Jev guesses this word, the game ends instantly.
Each turn, write a one-word clue that you think will hint at as many of the green cards as possible.
Click the cards you are trying to hint at (these aren't fed to Jev, but decide how many guesses can be made that turn).
If Jev guesses a non-key and non-bomb card (beige), the turn ends (no points awarded).
If the confidence is too low, there might also be fewer guesses made than allowed.
Get Jev to guess all 9 key cards to win. A goal of getting it in 6 clues is a nice challenge.
Why
Well, we get to see how Jev interprets your clue. A human-to-human game of Codenames involves subtlety and innuendo, knowledge of your partners, and creative stretching of ideas. How well can a model work with that kind of input? And also, how do you as a human change your clueing strategy to match the behavior of the model?
Who
An experiment by Aaron .
Close