Labeled positions for Texel tuning

[lithander]

I also noticed that it looks like the Leorik data is only decisive games - 1-0 or 0-1, no draws. I wonder if that affects the results, too.

If you want the bigger set with draws I can provide it but personally I don't use it: I just get better tuning results from ignoring all drawn games.

[Robert Pope]

No, I just noticed the difference. 3.5M positions is about the limit for what I can handle anyway.

[AndrewGrant]

I don't assert that the data I posted is high quality. It was some of the first data to be used for training of anything with ML approaches for the most part back in 2018 or maybe 2017. It was good enough that I was rapidly improving the HCE and closing the elo gap with Komodo (and now dishonored Houdini and Fire). I'm sure you can generate your own data and have better results with some effort.

[Robert Pope]

First test with tuning on Leorik data was a bust. -50 elo relative to my Zurichess tuning. However, I did just add a highly parameterized gamestage function, so I want to roll that back and see what that does first.