An amazing net for SF-NNUE.

[Sylwy]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

[Zenmastur]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

How does this net fair against some of Sergio's more recent nets?

[Laskos]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

Hmmm, there seem to be chunks of identical hex strings with Sergio nets. Strength-wise, also almost identical to latest Sergio nets.

[Eduard]

There are differences to the Sergio network. Here are some examples (german):

https://forum.computerschach.de/cgi-bin ... ?tid=11372

[Laskos]

There are differences to the Sergio network. Here are some examples (german):

https://forum.computerschach.de/cgi-bin ... ?tid=11372

These are a couple positions. On similarity Sim 8300 positions, this net is behaving as a late SV net among SV nets:

Code: Select all

sim version 3

  Key:

  1) Lc0 96x8 (time: 100 ms  scale: 1.0)
  2) Lc0 SV_3010 (time: 100 ms  scale: 1.0)
  3) Lc0 T40 (time: 100 ms  scale: 1.0)
  4) Lc0 T70 (time: 100 ms  scale: 1.0)
  5) SF_FiNN02 (time: 100 ms  scale: 1)
  6) SF_NNUE_GK (time: 100 ms  scale: 1)
  7) SF_NNUE_nightnurse (time: 100 ms  scale: 1)
  8) SF_NNUE_qhapaq (time: 100 ms  scale: 1)
  9) SF_NNUE_SV_1515 (time: 100 ms  scale: 1)
 10) SF_NNUE_SV_1515 Bis (time: 100 ms  scale: 1)
 11) SF_NNUE_SV_2141 (time: 100 ms  scale: 1)
 12) SF_NNUE_SV_2257 (time: 100 ms  scale: 1)
 13) Stockfish 10 (time: 100 ms  scale: 1)
 14) Stockfish 11 (time: 100 ms  scale: 1)
 15) Stockfish 6 (time: 100 ms  scale: 1)
 16) Stockfish 8 (time: 100 ms  scale: 1)
 17) Stockfish dev (time: 100 ms  scale: 1)
 18) Stockfish dev Bis (time: 100 ms  scale: 1)
 19) Andscacs (time: 100 ms  scale: 1)
 20) Komodo 14 (time: 100 ms  scale: 1)
 21) Ethereal (time: 100 ms  scale: 1)

         1     2     3     4     5     6     7     8     9    10    11    12    13    14    15    16    17    18    19    20    21
  1.  ----- 63.60 64.23 67.24 46.61 53.10 51.52 53.65 53.14 53.58 53.24 53.02 50.62 50.30 47.90 49.45 50.57 50.42 47.98 49.50 49.73
  2.  63.60 ----- 74.82 70.39 46.82 53.96 51.30 55.46 55.33 55.20 55.22 55.88 51.30 51.23 48.35 50.07 51.76 51.92 48.46 49.84 49.00
  3.  64.23 74.82 ----- 70.25 46.08 53.28 51.01 53.92 54.05 54.30 53.87 54.18 50.23 50.10 47.41 49.22 50.44 50.45 48.06 49.15 48.14
  4.  67.24 70.39 70.25 ----- 47.63 55.83 52.67 56.60 56.41 56.05 56.71 57.10 52.62 52.61 49.78 51.07 52.62 52.18 49.73 50.44 50.30
  5.  46.61 46.82 46.08 47.63 ----- 50.76 48.09 49.48 49.13 49.10 50.46 49.39 47.71 48.15 45.04 47.51 47.84 48.07 45.74 46.16 45.39
  6.  53.10 53.96 53.28 55.83 50.76 ----- 54.60 58.07 58.19 57.67 60.96 57.55 55.12 55.38 51.52 54.33 55.80 55.24 52.05 52.11 51.76
  7.  51.52 51.30 51.01 52.67 48.09 54.60 ----- 55.40 55.66 55.49 55.07 55.09 50.79 50.53 47.80 50.33 50.39 50.34 48.82 48.97 48.64
  8.  53.65 55.46 53.92 56.60 49.48 58.07 55.40 ----- 63.57 62.61 60.33 62.44 52.20 52.06 48.18 50.89 51.44 51.51 49.14 50.01 49.67
  9.  53.14 55.33 54.05 56.41 49.13 58.19 55.66 63.57 ----- 64.01 59.95 63.05 51.81 52.10 48.91 51.25 52.18 51.63 48.66 49.77 49.43
 10.  53.58 55.20 54.30 56.05 49.10 57.67 55.49 62.61 64.01 ----- 59.88 62.73 51.68 51.70 48.17 49.99 51.43 51.51 48.24 49.72 48.79
 11.  53.24 55.22 53.87 56.71 50.46 60.96 55.07 60.33 59.95 59.88 ----- 60.06 52.65 53.02 49.65 52.02 53.03 52.61 49.87 50.51 50.12
 12.  53.02 55.88 54.18 57.10 49.39 57.55 55.09 62.44 63.05 62.73 60.06 ----- 51.78 52.20 48.34 51.32 51.75 51.70 49.17 49.81 49.04
 13.  50.62 51.30 50.23 52.62 47.71 55.12 50.79 52.20 51.81 51.68 52.65 51.78 ----- 64.79 56.43 61.67 63.57 63.51 56.05 55.83 56.25
 14.  50.30 51.23 50.10 52.61 48.15 55.38 50.53 52.06 52.10 51.70 53.02 52.20 64.79 ----- 55.52 60.31 66.04 65.53 55.55 55.90 54.95
 15.  47.90 48.35 47.41 49.78 45.04 51.52 47.80 48.18 48.91 48.17 49.65 48.34 56.43 55.52 ----- 60.96 54.79 54.86 56.02 54.90 54.05
 16.  49.45 50.07 49.22 51.07 47.51 54.33 50.33 50.89 51.25 49.99 52.02 51.32 61.67 60.31 60.96 ----- 59.03 58.62 58.55 56.19 55.45
 17.  50.57 51.76 50.44 52.62 47.84 55.80 50.39 51.44 52.18 51.43 53.03 51.75 63.57 66.04 54.79 59.03 ----- 66.06 54.88 54.60 53.96
 18.  50.42 51.92 50.45 52.18 48.07 55.24 50.34 51.51 51.63 51.51 52.61 51.70 63.51 65.53 54.86 58.62 66.06 ----- 54.22 54.90 54.47
 19.  47.98 48.46 48.06 49.73 45.74 52.05 48.82 49.14 48.66 48.24 49.87 49.17 56.05 55.55 56.02 58.55 54.88 54.22 ----- 53.91 53.44
 20.  49.50 49.84 49.15 50.44 46.16 52.11 48.97 50.01 49.77 49.72 50.51 49.81 55.83 55.90 54.90 56.19 54.60 54.90 53.91 ----- 53.86
 21.  49.73 49.00 48.14 50.30 45.39 51.76 48.64 49.67 49.43 48.79 50.12 49.04 56.25 54.95 54.05 55.45 53.96 54.47 53.44 53.86 -----

Interesting also to note than self-similarity of SF dev on 4 threads (66%) is a bit higher than self-similarity of SF_NNUE_SV on 4 threads (64%).

[pohl4711]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

How does this net fair against some of Sergio's more recent nets?

I tested it vs. SF 200717, like all S.Vieri nets. Look here:
https://www.sp-cc.de/files/ergebnisse_nnue_schnell.txt

The net is very strong, but not the best in that tests. The best is still sv200730_1934

[corres]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

Is there another site to download it?
Dropbox is not a good place.

[tmokonen]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

Is there another site to download it?
Dropbox is not a good place.

There's nothing wrong with DropBox. Bizarre comment.

Also, any time there's a DropBox link that ends in ?dl=0, you can change the 0 to a 1 so that it's a direct download.

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=1

[Guenther]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

Hmmm, there seem to be chunks of identical hex strings with Sergio nets. Strength-wise, also almost identical to latest Sergio nets.

Good judgement...

As expected the cloning on NNUE nets has begun.
People can now test clones of engines with clones of nets w/o knowing they play A+A vs. A+A.

This so called qh-fine-0803-5 net is 99.99999% identical to Sergio Vieris net 20200729-0109.
Only at the very end of millions of hex entries 'someone' changed a few hex values ;-)

UltraEdit compare mode:

Code: Select all

Comparing: C:\Engines\NNUE-SV\qh-fine-0803-5.bin
To: C:\Engines\NNUE-SV\20200729-0109.bin
====
====
166883	 <! üúþúþû/öÿÿ=
	 !> üúþúþûAöÿÿ+
166893	 <! ùéööñÆØ°íÐúûúçûîëäÙ÷Ìýÿÿí\GÔ"Bñ&æ ë8ñ-#Ùîåà˼ðç
	 !> ùéööñÆØ°íÐúûúçûîëäÙ÷¸ýÿÿí\GÔ"Bñ&æ ë8ñ-#Ùîåà˼ðç

[corres]

Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0

Is there another site to download it?
Dropbox is not a good place.

There's nothing wrong with DropBox. Bizarre comment.
Also, any time there's a DropBox link that ends in ?dl=0, you can change the 0 to a 1 so that it's a direct download.
https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=1

Thanks for the advice. I did not know this trick. Is this bizarre?