[TODO] Add UCI column

#2
by christopher - opened

Avoids users having to parse the movetext column in case they need UCI

EDIT: this applies to all game datasets

That would be pretty awesome to have.

lichess.org org

Thank you for the feedback! I'll try to start working on this next week.

Awesome. I'm building a chess neural net on top of this data (working with parquet is way nicer than working with raw PGN). I ended up building a parser [1] to extract clock times as well; having clock data in a separate column in this dataset would help too.

[1] https://github.com/vladkvit/rust_pgn_reader_python_binding

lichess.org org

Very cool repo, thank you for sharing!

What would the clk column look like? I've been meaning to look into extracting clk and eval at some point. Are you planning to include clk in your model somehow?

Comma-separated integer milliseconds for clk would be nice. One downside I see with this approach is data storage size and unnecessary precision for 7+ day games.

I'm building a model that plays like a human (think MaiaChess + modernized architecture + several other features). It's currently achieving ~58% accuracy on move prediction vs Maia's ~52%.

I'm using clock data for two things. One is better move prediction, as moves under time pressure are less sophisticated and quicker. I'm also training a prediction head to predict how long a move takes. I am still actively reworking this, as move timing prediction is currently not predicting well. I haven't really dedicated much effort to this yet; still need to experiment with ideas like relative vs absolute time, log(clk), and clk / timecontrol_base.

Sign up or log in to comment