Closed sroelants closed 3 months ago
Fit a logistic WDL model to self-play data and rescale evals in the SF way (+1.0 eval means 50% chance of winning)
WDL percentages are only rendered when pretty printing, but evals/scores are rescaled in both pretty and uci output.
Generating data ought to be done through the wdl compile flag, which switches off the rescaling and outputs scores in the original, internal units.
wdl
I'll probably try and generate more, accurate data at some point (the pgns I used were from an SPSA, lol).
Fit a logistic WDL model to self-play data and rescale evals in the SF way (+1.0 eval means 50% chance of winning)
WDL percentages are only rendered when pretty printing, but evals/scores are rescaled in both pretty and uci output.
Generating data ought to be done through the
wdlcompile flag, which switches off the rescaling and outputs scores in the original, internal units.I'll probably try and generate more, accurate data at some point (the pgns I used were from an SPSA, lol).