Opening Test Suites releases

[chrisw]

https://github.com/ChrisWhittington/Chess-Openings-Data

Suitable for testers. Lines are active, usually with both sides having active play/counterplay. Selection criteria are highest non-draw result rate (W+L)/(W+D+L) from a large set of chess rating list computer-computer games. The idea being that no matter what the static evaluation of a line is, the important detail is the engine-engine result after playing the line. White bias is shown for each PGN, ideally we would like whitebias=0.5, but it is usually greater. Lines with whitebias > 0.7 are culled.
In practical play between strong engines I don't vouch for a much reduced draw rate, but I think you'll find the games are much more unbalanced in a play-counterplay sense. What an engine pair makes of that is up to the engine pair.

These PGNs:
require at least 50 examples each
are ranked on non-draw rate
use a whitebias filter of 0.7
use an eval filter of 200 (SF15 analysis at 1 core and 1 second)
are 11 ply deep

Complaints, criticisms, support, general comments - Use Github - Issues, that's what it's for.

[chrisw]

https://github.com/ChrisWhittington/Chess-Openings-Data

Suitable for testers. Lines are active, usually with both sides having active play/counterplay. Selection criteria are highest non-draw result rate (W+L)/(W+D+L) from a large set of chess rating list computer-computer games. The idea being that no matter what the static evaluation of a line is, the important detail is the engine-engine result after playing the line. White bias is shown for each PGN, ideally we would like whitebias=0.5, but it is usually greater. Lines with whitebias > 0.7 are culled.
In practical play between strong engines I don't vouch for a much reduced draw rate, but I think you'll find the games are much more unbalanced in a play-counterplay sense. What an engine pair makes of that is up to the engine pair.

These PGNs:
require at least 50 examples each
are ranked on non-draw rate
use a whitebias filter of 0.7
use an eval filter of 200 (SF15 analysis at 1 core and 1 second)
are 11 ply deep

Complaints, criticisms, support, general comments - Use Github - Issues, that's what it's for.

added two more opening suites, at depth 12

[gordonr]

Thanks for sharing these files

[bastiball]

https://github.com/ChrisWhittington/Chess-Openings-Data

Suitable for testers. Lines are active, usually with both sides having active play/counterplay. Selection criteria are highest non-draw result rate (W+L)/(W+D+L) from a large set of chess rating list computer-computer games. The idea being that no matter what the static evaluation of a line is, the important detail is the engine-engine result after playing the line. White bias is shown for each PGN, ideally we would like whitebias=0.5, but it is usually greater. Lines with whitebias > 0.7 are culled.
In practical play between strong engines I don't vouch for a much reduced draw rate, but I think you'll find the games are much more unbalanced in a play-counterplay sense. What an engine pair makes of that is up to the engine pair.

These PGNs:
require at least 50 examples each
are ranked on non-draw rate
use a whitebias filter of 0.7
use an eval filter of 200 (SF15 analysis at 1 core and 1 second)
are 11 ply deep

Complaints, criticisms, support, general comments - Use Github - Issues, that's what it's for.

Thanks a lot!

[chrisw]

For Testers and Testing groups:

On Github is a 100 line opening suite which is same as before, except it's now filtered to remove lines with an SF15 eval < 50 cp.

low-drawrate-openings-v2-q50-n100-d11-Wbiasfilter0.7-Evalfilter50-100-comp-comp.pgn

Tentative results (at TC=20+0.05, so very fast games) against my usual pool of top engines are:

Using random book, drawrate = 0.60
Using UHO book, drawrate = 0.50
Using above book, drawrate = 0.48

which indicates an improvement. Obviously nowhere near enough data, but looks like this one may be worth testing ...
If it looks any good, I'll try and generate a 1000-liner as well.

Comments welcome