CCRL pgn database - what pov do the included scores have?

[Desperado]

Hello everybody.

I am curious. The games I looked at manually seem to contain the evaluations from White's point of view.

Can anyone confirm this and is it consistent across all games?
Since different testers work with different tools, I can imagine that different GUIs save the games (the scores) differently.

I am alternatively thinking about the Lichess database, which always uses "wpov". However, these are "only" Stockfish scores.

Does anyone know of a database with several million computer games that consistently contains the evaluation from the point of view of the side on the move or from White's point of view?

Many thanks for any feedback.

[Desperado]

Hello again.

As I have unfortunately not yet received an answer, I will try to go into more detail about why I want to know this.

It is about the following link https://computerchess.org.uk/ccrl/4040/games.html and the data that is commented on.

I have a tuner that can parse the data, i.e. also reads out the scores. As there are almost two million games involved, I can't check every game manually, of course. Of course, the tuner only works if the perspective of the evaluations is the same for all games, or at least the perspective from which the evaluation is saved can be determined automatically. This should therefore always be WPOV, or always from the perspective of the player making the move.

Thank you.

[dkappe]

A couple of years ago, the lc0 project massaged the CCRL and CEGT pgn’s for training. Have a google for that info.

[Desperado]

A couple of years ago, the lc0 project massaged the CCRL and CEGT pgn’s for training. Have a google for that info.

Thanks for the hint. I was able to find https://lczero.org/blog/2018/09/a-standard-dataset/ this blog.
They used the data, but unfortunately not necessarily the evaluations.(The downloads no longer seem to work either.)
I can't find any information on my specific question either.

Now, I have tons of data to feed the tuner. However, the signals are then always based on the result of the games and not the move evaluations.
I can also generate games myself (with different engines), but not 15 minute games with a comparable number of games.

Who are the testers of CCRL, they should really know it. Unfortunately, I have not been able to find any information in the forum or on the CCRL website. (I may have simply overlooked it).

Thank you very much, I hope someone comes forward who can answer this question clearly.

Edit:
I found the list of contributors at CCRL. I think some of them are (still) active here. Maybe I need a little more patience

[Graham Banks]

I'll let Ray answer this.
I would imagine from a white POV, but I used to do my testing with the engine POV.
Whether or not they all automatically conform to a white POV, I'm unsure.

[Modern Times]

I downloaded some of the games. Some have evals from white's pov, some have them from the engine's pov.

[Modern Times]

White pov

[Site "CCRL"]
[Date "2009.10.08"]
[Round "196.1.733"]
[White "Fritz 12"]
[Black "Onno 1.1.1 32-bit"]
[Result "1-0"]
[ECO "B19"]
[Opening "Caro-Kann"]
[Variation "classical, Spassky variation"]
[PlyCount "109"]

<snip>
53. Rd3 {(Rh1) +6.67/16 16s} f3 {(Qg6) +6.16/16 45s}
54. Qxf3 {(Qxc6) +7.77/16 7s} Qg1+ {+5.74/15 15s}
55. Rd1 {+8.29/17 9s} 1-0


Engine pov

[Event "CCRL 40/15"]
[Site "CCRL"]
[Date "2024.01.10"]
[Round "914.9.294"]
[White "Seer 2.8.0 64-bit"]
[Black "RubiChess 20230918 64-bit"]
[Result "1-0"]
[ECO "B30"]
[Opening "Sicilian"]
[Variation "Nimzovich-Rossolimo attack (without ...d6)"]
[PlyCount "222"]

<snip>
107. Kf3 {(Kf3) +6.65/31 8s} Bc1 {(Bc1) -4.51/29 8s} 108. Nd3 {(h6) +7.14/30 5s} Bg5 {(Bd2) -5.05/36 8s}
109. Ke2 {(Nc5) +7.78/24 4s} Bh4 {(Ke8) -5.11/38 7s} 110. f4 {(Ke3) +9.25/28 5s} Kf8 {(Kf8) -6.28/31 8s}
111. Kf3 {(Kf3) +10.09/30 7s} Ke8 {(Ke8) -7.42/29 8s} 1-0

[Desperado]

Thank you for taking care of it.

In the meantime, I was also able to do the same. Unfortunately, the perspectives for the evaluations are not uniform.

I will now think about whether there is a way to automatically recognize the perspective of the evaluations. You "only" need one unique piece of information per game to be certain. For decided games, the principle is simple, by looking at the evaluations of the last moves, their deviations from each other and from the result ... For games that end in a draw, this is not immediately obvious (at least not to me).

If I am able to find a good heuristic, I can convert the scores for my purposes.
I'm already working on my idea, but if anyone already knows a solution to this issue, I'm very curious.
Finally a game that ended in a draw which seem to have engine pov.

Code: Select all

[Event "CCRL 40/15"]
[Site "CCRL"]
[Date "2020.12.12"]
[Round "749.1.214"]
[White "Orion 0.8 64-bit"]
[Black "Tucano 9.00 64-bit"]
[Result "1/2-1/2"]
[ECO "A04"]
[Opening "Reti"]
[Variation "Lisitsin gambit deferred"]
[PlyCount "104"]
[WhiteElo "3112"]
[BlackElo "2939"]

1. Nf3 {+0.00/1 0s} f5 {+0.00/1 0s} 2. d3 {+0.00/1 0s} Nf6 {+0.00/1 0s} 3. e4
{+0.00/1 0s} d6 {+0.00/1 0s} 4. Be2 {+0.00/1 0s} e5 {+0.00/1 0s} 5. O-O
{+0.00/1 0s} Be7 {+0.00/1 0s} 6. Nc3 {+0.00/1 0s} O-O {+0.00/1 0s} 7. a4
{+0.00/1 0s} c5 {+0.00/1 0s} 8. Nd5 {+0.00/1 0s} Nxd5 {+0.00/1 0s} 9. exd5
{-0.44/18 31s} Na6 {(Na6) +0.44/24 80s} 10. c3 {(c3) -0.35/16 25s} Nc7 {(Nc7)
+0.46/25 70s} 11. a5 {(Re1) -0.31/17 24s} Bf6 {(Bf6) +0.41/24 50s} 12. c4
{(Qb3) -0.32/16 48s} Na6 {(Kh8) +0.32/25 88s} 13. Ne1 {(Ne1) -0.39/17 27s} e4
{(Bd7) +0.34/25 60s} 14. Nc2 {(Qb3) -0.11/17 32s} Be5 {(Be5) +0.10/26 153s} 15.
Ra3 {(Ra3) -0.02/17 50s} exd3 {(Qh4) +0.05/23 84s} 16. Bxd3 {(Qxd3) +0.11/17
35s} Qc7 {(f4) +0.00/24 48s} 17. Ne1 {(Re1) +0.29/15 40s} Nb4 {(Nb4) -0.33/26
115s} 18. Bb1 {(Bb1) +0.51/17 51s} Bf6 {(Bf6) -0.39/23 27s} 19. Nf3 {(Qb3)
+0.60/16 33s} h6 {(h6) -0.27/23 27s} 20. Ne1 {(Ne1) +0.34/16 45s} Rb8 {(Qe7)
-0.35/22 26s} 21. Rg3 {(Nd3) +0.37/17 52s} Qxa5 {(Be5) -0.29/25 79s} 22. Bxh6
{(Bxh6) +1.33/15 28s} b5 {(Qa6) -0.77/24 56s} 23. Qh5 {(Bf4) +1.28/16 38s} bxc4
{(bxc4) -1.11/23 62s} 24. Nf3 {(Nf3) +1.43/16 40s} Rb7 {(Rb7) -0.88/24 33s} 25.
Bd2 {(Ng5) +2.01/16 46s} Qa4 {(Qd8) -1.40/23 60s} 26. Rh3 {(Rh3) +1.47/16 41s}
g6 {(g6) -1.37/23 30s} 27. Qxg6+ {(Qxg6) +1.70/17 47s} Rg7 {(Rg7) -1.09/24 40s}
28. Qh6 {(Qh6) +1.70/18 39s} Qe8 {(Qe8) -1.62/25 33s} 29. Nh4 {(Nh4) +1.60/17
59s} Nd3 {(Nd3) -1.72/24 24s} 30. Bxd3 {(Bxd3) +1.29/17 46s} cxd3 {(cxd3)
-1.45/26 72s} 31. Re1 {(Re1) +1.63/18 33s} Qd8 {(Qd8) -1.37/24 20s} 32. Nf3
{(Ng6) +1.80/18 33s} f4 {(f4) -1.87/24 38s} 33. Re6 {(Re6) +1.48/19 54s} Bxe6
{(Bxe6) -1.95/24 34s} 34. dxe6 {(dxe6) +1.60/20 69s} Rb7 {(Rb7) -1.92/23 24s}
35. Qxf4 {(Qxf4) +1.49/17 55s} d5 {(Bxb2) -1.49/22 31s} 36. Qg4+ {(Qg4)
+2.35/18 44s} Bg7 {(Bg7) -0.97/20 7s} 37. Qg6 {(Qg6) +2.13/17 79s} Rf6 {(Rf6)
-0.98/20 6s} 38. Qxd3 {(Qh7) +1.93/17 79s} Qe8 {(Qe8) -1.06/18 3s} 39. Qxd5
{(Qh7) +1.66/16 45s} Rxb2 {(Rxb2) -1.03/20 1s} 40. Qd3 {(Qd3) +1.04/19 71s}
Rxe6 {(Qxe6) -0.45/17 0s} 41. Qh7+ {(Qh7) +1.88/21 38s} Kf8 {(Kf8) +0.00/24
63s} 42. Rg3 {(Be3) +2.75/20 20s} Qf7 {(Qf7) +0.00/27 50s} 43. Rxg7 {(Rxg7)
+2.54/20 19s} Qxg7 {(Qxg7) +0.00/28 185s} 44. Qxg7+ {(Qxg7) +1.62/22 29s} Kxg7
{(Kxg7) +0.00/32 59s} 45. Bc3+ {(Bc3) +1.99/21 36s} Kg6 {(Kh7) +0.00/34 152s}
46. Bxb2 {(Bxb2) +1.95/20 32s} Rb6 {(Rb6) +0.00/34 132s} 47. Bc3 {(Bc3)
+1.98/19 23s} Rb1+ {(Rb3) +0.00/32 45s} 48. Ne1 {(Ne1) +0.00/23 27s} Rc1 {(Rc1)
+0.00/45 35s} 49. Bd2 {(Bd2) +0.00/25 35s} Rd1 {(Rd1) +0.00/46 36s} 50. Bc3
{(Ba5) +0.00/26 29s} Rc1 {(Rc1) +0.00/55 36s} 51. Bd2 {+0.00/26 41s} Rd1 {(Rd1)
+0.00/57 45s} 52. Bc3 {(Bc1) +0.00/27 41s} Rc1 {(Rc1) +0.00/52 62s} 1/2-1/2

Maybe at some point this thread should be moved to the programming section. Let's see.

Code: Select all

31. Re1 {(Re1) +1.63/18 33s} Qd8 {(Qd8) -1.37/24 20s
35. Qxf4 {(Qxf4) +1.49/17 55s} d5 {(Bxb2) -1.49/22 31s}

This information indicates an engine pov. However, this could theoretically also be wpov,
where the respective engine disagree in evaluation.