Repository navigation
Replies: 2 comments
|
Thanks for asking! I haven't dug very far into this yet, but I took a quick look and found a few things worth sharing. Some background first: for every game, each language model gets a randomly chosen setup. That means SAN or UCI notation, the FEN and/or the move history, a board diagram or not, the list of legal moves or not, and one of several prompt templates. So we can compare setups fairly. I looked at about 32,000 games and 30,000 puzzle attempts, mostly at how often models play illegal moves (which forfeits the game here), plus game results and puzzle solve rates. Here's what stood out:
This was only a quick pass. I haven't looked closely at move quality, at how these settings combine, or at models with only a few games. If you're curious about a specific model, or anything else, let me know and I can take a closer look! |
|
Cool, thanks for the info |
Uh oh!
There was an error while loading. Please reload this page.
Is there anywhere I can see which move notation, position info and prompt works best for each model?
All reactions