For data lovers only

[jorose]

So far I have 3.7 million human-human games between at least 2200 elo rated players.

Why was 2200 elo chosen as the cutting point? I believe it's equivalent to 1900 elo of chess.com.

That might have been true at some point in the passed, but nowadays the distributions match fairly closely in that rating range. 2200 Lichess blitz is roughly 2200 chess.com blitz.

[Rebel]

All done...

In total 25.6 million annotated Stockfish games.

Say average game length is 60 moves, we get:

25.000.000 x 120 = 3.000.000.000 positions

for NN training or other purposes.

[fersbery]

Thanks for this, it will be very useful for supervised learning.

OT: Long time fan of Rebel here, remember playing a lot of games vs one of the old DOS versions.

[jorose]

All done...

In total 25.6 million annotated Stockfish games.

Say average game length is 60 moves, we get:

25.000.000 x 120 = 3.000.000.000 positions

for NN training or other purposes.

Are they SF games (... against SF?) annotated by SF or are they just arbitrary games annotated by SF?

[Dann Corbit]

All done...

In total 25.6 million annotated Stockfish games.

Say average game length is 60 moves, we get:

25.000.000 x 120 = 3.000.000.000 positions

for NN training or other purposes.

If you post a link, I will be grateful to greedily snatch a copy.

[brianr]

Are the links not the ones in the first post?

[Rebel]

All done...

In total 25.6 million annotated Stockfish games.

Say average game length is 60 moves, we get:

25.000.000 x 120 = 3.000.000.000 positions

for NN training or other purposes.

If you post a link, I will be grateful to greedily snatch a copy.

It's on the link in the OP - http://rebel13.nl/download/data.html

[Ovyron]

That might have been true at some point in the passed, but nowadays the distributions match fairly closely in that rating range. 2200 Lichess blitz is roughly 2200 chess.com blitz.

When analyzing the quality of the moves people play, it seems chess.com's 1300 elo players play at the same level as 1600 elo players of Lichess. I wonder if there's a tool to check for this automatically. But, yeah, my mistake was assuming this was an offset in rating, I didn't expect the difference to shrink the higher the rating so that lichess "catches up" at some point.

[Dann Corbit]

All done...

In total 25.6 million annotated Stockfish games.

Say average game length is 60 moves, we get:

25.000.000 x 120 = 3.000.000.000 positions

for NN training or other purposes.

If you post a link, I will be grateful to greedily snatch a copy.

It's on the link in the OP - http://rebel13.nl/download/data.html

Too bad there are no search depth indicators in the data.
It is hard to know what time of search means since we don't know the hardware so the evaluations can only be taken as a very crude measurement.

[jorose]

That might have been true at some point in the passed, but nowadays the distributions match fairly closely in that rating range. 2200 Lichess blitz is roughly 2200 chess.com blitz.

When analyzing the quality of the moves people play, it seems chess.com's 1300 elo players play at the same level as 1600 elo players of Lichess. I wonder if there's a tool to check for this automatically. But, yeah, my mistake was assuming this was an offset in rating, I didn't expect the difference to shrink the higher the rating so that lichess "catches up" at some point.

It's hard to say exactly. This link gives a rough impression and is based on player surveys. I imagine the data has some pretty big error bars though. My understanding is the gap used to be larger between the two sites, but chess.com ratings have inflated quite a bit in recent years.