A new Benjamin ?

Benjamin 1.0 was the gambit version of ProDeo.

I am now working on a possible Benjamin 2.0 version meant to be the gambit version of Rebel.

Elo is not the end to the means, playing style is, and Stefan's EAS tool is still the best around to measure aggressiveness. This will be a project for months generating data with occasional updates.

Code:

 Net size 850M   | Net size 1.7B  |
 Epoch     EAS   | Epoch   EAS    |
   8     267044  |  16   334680   |
  16     234491  |  32   282878   |
                 |  39   280698   |

Currently the 1.7B version shows a loss of 140 elo in comparison to Rebel 16.3, I have hope the loss will be limited to ~100 elo when I have generated 10B of data, estimated hopefully still in 2024  

Surprisingly I noticed the new generated data (so far) doesn't bite the large 25.5B Rebel 16.3 data, it even gave a small improvement and also a higher EAS score. But there are a lot of maybe's here.

All in all I have the fun back in NNUE programming after the discord debacle. I grew up in the 80's with originality, originality and originality and I prefer to stick with that, but maybe I am old fashioned.

[pgn][Event "?"]
[Site "?"]
[Date "2024.09.17"]
[Round "78"]
[White "RubiChess-20230410"]
[Black "Benjamin-E16"]
[Result "0-1"]
[ECO "E97"]
[Opening "King's Indian"]
[Variation "Orthodox, Aronin-Taimanov Variation (Yugoslav attack / Mar del Plata Variation )"]
[TimeControl "40/10"]
[PlyCount "64"]

1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5 7. O-O Nc6 8.
d5 Ne7 9. Nd2 c6 10. a3 Ne8 11. b4 f5 12. Qb3 Kh8 13. a4 f4 14. Bb2 g5 15.
h3 Ng6 16. a5 g4 17. Bxg4 Bxg4 18. hxg4 f3 19. Nxf3 Rxf3 20. a6 Rb8 21.
axb7 Nf4 22. gxf3 Qh4 23. Bc1 Bh6 24. Bxf4 Bxf4 25. Rfb1 Qh3 26. Rb2 Nf6
27. dxc6 Bh2+ 28. Kh1 Qxf3+ 29. Kxh2 Nxg4+ 30. Kg1 Ne3 31. fxe3 Rg8+ 32.
Rg2 Qxg2# [/pgn]

Nice shorty in the style of the King's Indian.

Admin wrote:

Benjamin 1.0 was the gambit version of ProDeo.

I am now working on a possible Benjamin 2.0 version meant to be the gambit version of Rebel.

Elo is not the end to the means, playing style is, and Stefan's EAS tool is still the best around to measure aggressiveness. This will be a project for months generating data with occasional updates.

Code:

 Net size 850M   | Net size 1.7B  |
 Epoch     EAS   | Epoch   EAS    |
   8     267044  |  16   334680   |
  16     234491  |  32   282878   |
                 |  39   280698   |

EAS scores of 280k-334k are very impressive! Of course, such high EAS scores must cost Celos. Playing high risky style is nice - but you have to pay a price... Celos have to be invested - no way to avoid this. Only question is, how many Celos. IMHO a Celo-investment below -150 is already a good trade.

I keep my fingers crossed! Cant wait to see a new Rebel EAS!!!

Fun...

Code:

Net size 850M   | Net size 1.7B   |
 Epoch     EAS   | Epoch   EAS    |
   8     267044  |  16   334680   |
  16     234491  |  32   282878   |
                 |  39   280698   |

I have been playing a 3000+ game match between Patricia and the epoch 39 version, the EAS :

Code:

Rank  EAS-Score  sacs   shorts  draws  moves  Engine/player
-------------------------------------------------------------------
   1    118483  21.23%  10.96%  12.29%   70   patricia-3.01  
   2     47564  03.84%  17.99%  28.79%   64   Benjamin-E39  
-------------------------------------------------------------------

Gone is Patricia's 400K score
Gone is Benjamin's 300K score

Oh, then the match score :

Code:

No. Name           Win Draw Loss Unf.  Score Games       %
----------------------------------------------------------
  1 Benjamin-E39  +1823 =1115 -146   *0 2380.5  3084   77.2%
  2 patricia-3.01 +146 =1115 -1823   *0  703.5  3084   22.8%

Ouch, I thought Patricia was a lot stronger.

UPDATE - Added 500M to the Benjamin data, now 2.2B

Code:

 Net size 850M   | Net size 1.7B  | Net size 2.2B
 Epoch     EAS   | Epoch   EAS    | Epoch   EAS
   8     267044  |  16   334680   |  22   330782
  16     234491  |  32   282878   |  39   291125
                 |  39   280698   |  59   380855
                                  |  66   295026

Next update after 3B.

Admin wrote:

Fun...

Code:

Net size 850M   | Net size 1.7B   |
 Epoch     EAS   | Epoch   EAS    |
   8     267044  |  16   334680   |
  16     234491  |  32   282878   |
                 |  39   280698   |

I have been playing a 3000+ game match between Patricia and the epoch 39 version, the EAS :

Code:

Rank  EAS-Score  sacs   shorts  draws  moves  Engine/player
-------------------------------------------------------------------
   1    118483  21.23%  10.96%  12.29%   70   patricia-3.01  
   2     47564  03.84%  17.99%  28.79%   64   Benjamin-E39  
-------------------------------------------------------------------

Gone is Patricia's 400K score
Gone is Benjamin's 300K score

Oh, then the match score :

Code:

No. Name           Win Draw Loss Unf.  Score Games       %
----------------------------------------------------------
  1 Benjamin-E39  +1823 =1115 -146   *0 2380.5  3084   77.2%
  2 patricia-3.01 +146 =1115 -1823   *0  703.5  3084   22.8%

Ouch, I thought Patricia was a lot stronger.

This result has nothing to do with the lower strength of Patricia. This is (for me) a well known effect: When 2 very aggressive playing engines just play each other, the EAS-scores are much smaller for both engines, compared to the EAS-scores of these engines, when playing vs a pool of "normal" playing engines. And this is not surprising, IMHO, because both steal EAS from the opponent:
A very aggressive engines plays a lot of speculative sacs, which often go wrong and lead to a loss. But in these games, the (aggressive) opponent has no chance, to play an own sac. So, the EAS-scores are much smaller for both engines as a result (lost games with a sac from the loosing color, are not counted by the EAS-tool)
You should play 2 gauntlets (of Patricia and Benjamin) versus a pool of normal engines and compare the EAS-results of these 2 gauntlet-testruns...

Yeah, they bite each other, where two fight for EAS both got damaged, fun to watch, no need to repeat the experiment.

Started a new training session for a possible REBEL EAS 2.0

Trainer states - Training positions: 27272727272

But I am not superstitious

[pgn][Event "?"]
[Site "?"]
[Date "2024.09.26"]
[Round "149"]
[White "mp-ben-E252"]
[Black "Seer-2.6.0"]
[Result "1-0"]
[ECO "B43"]
[Opening "Sicilian"]
[Variation "Kan, 5.Nc3"]
[TimeControl "40/10"]
[Termination "adjudication"]
[PlyCount "72"]

1. e4 c5 2. Nf3 e6 3. d4 cxd4 4. Nxd4 a6 5. Nc3 b5 6. Bd3 Qb6 7. Nf3 Nc6 8.
O-O Qb8 9. a4 b4 10. Nd5{*} exd5 11. exd5 Nd8 12. Re1+ Be7 13. Nd4 d6 14. Bf4
Kf8 15. Qf3 h5 16. h3 Qa7 17. Be3 Qb7 18. c4 bxc3 19. bxc3 Bf6 20. c4 g6
21. Rab1 Qc7 22. c5 Kg7 23. Nb3 h4 24. c6 Ne7 25. Nd2 Qa5 26. c7 Qxc7 27.
Qxf6+{*} Kxf6 28. Bd4+ Kg5 29. Bg7{*} Nxd5 30. Ne4+ Kf4 31. g3+ hxg3 32. fxg3+
Kf3 33. Bd4 Bxh3 34. Rf1+ Bxf1 35. Rxf1+ Kg4 36. Be2+ Kh3 1-0[/pgn]

Nice sac game, 10.Nd5!, 26.c7!, 27.Qxf6+, 29.Bg7!

i sincerely think adding human data helps, and makes engines stronger. for example, when i added some net trained on only human games to stockfish classical, it makes it stronger, maybe filling gaps. human games online are very sharp and provide much training data to help.

for example, here's a leela net at only 10k steps trained on human games, that plays more aggressive than any aggressive engine at its elo, while being ridiculously sharp for leela.
[pgn][Event "new nets test"]
[Site "Lucas Chess R 2.14l"]
[Date "2024.09.25"]
[Round "?"]
[White "Tao 5.6"]
[Black "Leela Tal Blitz 10k steps"]
[Result "0-1"]
[WhiteElo "2491"]
[BlackElo "2300"]
[ECO "A00"]
[Opening "Van Geet Opening: Napoleon Attack"]
[TimeControl "12+0"]
[Termination "Mate"]
[PlyCount "120"]
[Link "https://www.chess.com/analysis/game/pgn/26kfbJDzpW?tab=analysis&move=47"]

1. Nc3 e5 2. Nf3 Nc6 3. d4 exd4 4. Nxd4 Nf6 5. Bg5 Bb4 6. Qd3 h6 7. Qe3+ Kf8 8.
Bxf6 $6 Qxf6 9. Nb5 d5 $2 10. Nxc7 $6 Be6 11. Nxa8 $2 d4 $1 12. Qd3 dxc3 13. b3 Ke7 $2 14.
Qb5 Rd8 15. Qxb7+ $2 Kf8 $2 16. e3 Bc8 17. Qb5 Rd2 18. Be2 Bg4 19. f3 Qe7 $2 20. fxg4 $2
Qxe3 21. Nc7 $6 Kg8 $2 22. Nd5 $1 Rxd5 23. Qxc6 Rd4 $2 24. Qc8+ Kh7 25. Rf1 Rd5 26.
Qb7 $4 Qd2+ $1 27. Kf2 Re5 $2 28. Kg3 $2 Rxe2 29. Kh3 $2 f5 $9 30. Rxf5 Be7 31. Raf1 Qxc2
32. Qxa7 $2 Qe4 $1 33. R5f3 c2 34. Qa5 Bg5 $6 35. Qc5 $6 Rd2 36. Re1 $1 Qg6 37. Rc3 $2 Rd1
38. Qg1 $9 Bd2 39. Rxd1 Bxc3 40. Rc1 h5 $1 41. g5 Bd2 42. g3 $6 Qe6+ 43. Kg2 Qe2+
44. Kh3 Bxc1 45. Qxc1 Kg6 46. Qf4 Qg4+ 47. Qxg4 hxg4+ 48. Kxg4 c1=Q 49. h4 Qe3
50. h5+ Kf7 51. h6 gxh6 52. gxh6 $6 Qxh6 $6 53. Kf5 Qe3 54. g4 Ke7 55. Kg6 $6 Qf4
56. g5 Ke6 57. Kh6 Kf7 58. Kh5 Qg3 59. g6+ Kf6 60. b4 Qg5# 0-1

[/pgn]
what it learned was to play for mate, organizing its pieces and going for the win. it feels like if tal played online blitz today. i've had it run on lichess, and it equalized a completely lost position, but lost in the endgame.


[pgn][Event "Rated blitz game"]
[Site "lichess.org"]
[Date "2024.09.25"]
[Round "?"]
[White "Nakshatra3"]
[Black "EliteExperimental"]
[Result "0-1"]
[ECO "A48"]
[WhiteElo "2142"]
[BlackElo "2207"]
[PlyCount "126"]
[Beauty "6188376854094"]
[EventDate "2024.??.??"]
[EventType "blitz"]
[TimeControl "180+1"]

1. d4 Nf6 2. Nf3 g6 3. Bf4 Bg7 4. e3 O-O 5. Nbd2 d6 6. h3 Nbd7 7. Be2 Re8 8. O-O e5 9. Bg3 e4 10. Ne1 h5 11. c4 Bh6 12. a4 Nf8 13. a5 N8h7 14. Nb1 Ng5 15. d5 Bxh3 16. gxh3 Nxh3+ 17. Kh1 Qd7 18. a6 bxa6 19. Rxa6 Ng5 20. Kg1 Rab8 21. Qc2 Qh3 22. Nd2 h4 23. Bf4 Ng4 24. Rxa7 Ne6 25. dxe6 Bxf4 26. Nef3 Rxe6 27. exf4 exf3 28. Nxf3 Re3 29. Re1 Rxf3 30. Bxf3 Qxf3 31. Ra3 Qxf4 32. Rh3 g5 33. b3 Ne5 34. Re4 Qf5 35. Kg2 Ra8 36. Qe2 Kg7 37. b4 Ra1 38. b5 f6 39. f3 Ng6 40. Rh1 Ra8 41. Qd2 Ra3 42. Re3 Nf4+ 43. Kf1 Qb1+ 44. Qe1 Qc2 45. Re7+ Kh6 46. Rxh4+ gxh4 47. Qxh4+ Nh5 48. Qf2 Qxc4+ 49. Kg1 Ra1+ 50. Re1 Rxe1+ 51. Qxe1 Qxb5 52. Qe7 Qg5+ 53. Kf1 Qf4 54. Ke2 Qe5+ 55. Qxe5 fxe5 56. Kd1 d5 57. Kc1 e4 58. Kb1 exf3 59. Ka1 f2 60. Kb1 f1=Q+ 61. Ka2 Qb1+ 62. Ka3 c5 63. Ka4 Qb4# {Normal} 0-1

[/pgn]

And this game, is just beautiful.

texium wrote:

i sincerely think adding human data helps, and makes engines stronger. for example, when i added some net trained on only human games to stockfish classical, it makes it stronger, maybe filling gaps. human games online are very sharp and provide much training data to help.

for example, here's a leela net at only 10k steps trained on human games, that plays more aggressive than any aggressive engine at its elo, while being ridiculously sharp for leela.
[pgn][Event "new nets test"]
[Site "Lucas Chess R 2.14l"]
[Date "2024.09.25"]
[Round "?"]
[White "Tao 5.6"]
[Black "Leela Tal Blitz 10k steps"]
[Result "0-1"]
[WhiteElo "2491"]
[BlackElo "2300"]
[ECO "A00"]
[Opening "Van Geet Opening: Napoleon Attack"]
[TimeControl "12+0"]
[Termination "Mate"]
[PlyCount "120"]
[Link "https://www.chess.com/analysis/game/pgn/26kfbJDzpW?tab=analysis&move=47"]

1. Nc3 e5 2. Nf3 Nc6 3. d4 exd4 4. Nxd4 Nf6 5. Bg5 Bb4 6. Qd3 h6 7. Qe3+ Kf8 8.
Bxf6 $6 Qxf6 9. Nb5 d5 $2 10. Nxc7 $6 Be6 11. Nxa8 $2 d4 $1 12. Qd3 dxc3 13. b3 Ke7 $2 14.
Qb5 Rd8 15. Qxb7+ $2 Kf8 $2 16. e3 Bc8 17. Qb5 Rd2 18. Be2 Bg4 19. f3 Qe7 $2 20. fxg4 $2
Qxe3 21. Nc7 $6 Kg8 $2 22. Nd5 $1 Rxd5 23. Qxc6 Rd4 $2 24. Qc8+ Kh7 25. Rf1 Rd5 26.
Qb7 $4 Qd2+ $1 27. Kf2 Re5 $2 28. Kg3 $2 Rxe2 29. Kh3 $2 f5 $9 30. Rxf5 Be7 31. Raf1 Qxc2
32. Qxa7 $2 Qe4 $1 33. R5f3 c2 34. Qa5 Bg5 $6 35. Qc5 $6 Rd2 36. Re1 $1 Qg6 37. Rc3 $2 Rd1
38. Qg1 $9 Bd2 39. Rxd1 Bxc3 40. Rc1 h5 $1 41. g5 Bd2 42. g3 $6 Qe6+ 43. Kg2 Qe2+
44. Kh3 Bxc1 45. Qxc1 Kg6 46. Qf4 Qg4+ 47. Qxg4 hxg4+ 48. Kxg4 c1=Q 49. h4 Qe3
50. h5+ Kf7 51. h6 gxh6 52. gxh6 $6 Qxh6 $6 53. Kf5 Qe3 54. g4 Ke7 55. Kg6 $6 Qf4
56. g5 Ke6 57. Kh6 Kf7 58. Kh5 Qg3 59. g6+ Kf6 60. b4 Qg5# 0-1

[/pgn]
what it learned was to play for mate, organizing its pieces and going for the win. it feels like if tal played online blitz today. i've had it run on lichess, and it equalized a completely lost position, but lost in the endgame.


[pgn][Event "Rated blitz game"]
[Site "lichess.org"]
[Date "2024.09.25"]
[Round "?"]
[White "Nakshatra3"]
[Black "EliteExperimental"]
[Result "0-1"]
[ECO "A48"]
[WhiteElo "2142"]
[BlackElo "2207"]
[PlyCount "126"]
[Beauty "6188376854094"]
[EventDate "2024.??.??"]
[EventType "blitz"]
[TimeControl "180+1"]

1. d4 Nf6 2. Nf3 g6 3. Bf4 Bg7 4. e3 O-O 5. Nbd2 d6 6. h3 Nbd7 7. Be2 Re8 8. O-O e5 9. Bg3 e4 10. Ne1 h5 11. c4 Bh6 12. a4 Nf8 13. a5 N8h7 14. Nb1 Ng5 15. d5 Bxh3 16. gxh3 Nxh3+ 17. Kh1 Qd7 18. a6 bxa6 19. Rxa6 Ng5 20. Kg1 Rab8 21. Qc2 Qh3 22. Nd2 h4 23. Bf4 Ng4 24. Rxa7 Ne6 25. dxe6 Bxf4 26. Nef3 Rxe6 27. exf4 exf3 28. Nxf3 Re3 29. Re1 Rxf3 30. Bxf3 Qxf3 31. Ra3 Qxf4 32. Rh3 g5 33. b3 Ne5 34. Re4 Qf5 35. Kg2 Ra8 36. Qe2 Kg7 37. b4 Ra1 38. b5 f6 39. f3 Ng6 40. Rh1 Ra8 41. Qd2 Ra3 42. Re3 Nf4+ 43. Kf1 Qb1+ 44. Qe1 Qc2 45. Re7+ Kh6 46. Rxh4+ gxh4 47. Qxh4+ Nh5 48. Qf2 Qxc4+ 49. Kg1 Ra1+ 50. Re1 Rxe1+ 51. Qxe1 Qxb5 52. Qe7 Qg5+ 53. Kf1 Qf4 54. Ke2 Qe5+ 55. Qxe5 fxe5 56. Kd1 d5 57. Kc1 e4 58. Kb1 exf3 59. Ka1 f2 60. Kb1 f1=Q+ 61. Ka2 Qb1+ 62. Ka3 c5 63. Ka4 Qb4# {Normal} 0-1

[/pgn]

And this game, is just beautiful.

Could you point me to where I can find this interesting Leela Tal 10k steps net?

that was a bad pool, filled with aggressive engines, which i know lowers eas rating. here i tested using engines from ccrl, assuming 26-2700, and ccrl testing standards.
i did 20 games per engine. 10 on each color, ccrl doesnt even do that many. it only had 18 draws out of 120 games. though with 46 wins, i doubt it would get much lower with a larger sample

Brendan wrote:

texium wrote:

And this game, is just beautiful.

Could you point me to where I can find this interesting Leela Tal 10k steps net?

i can send it to you privately.

I too am interested. Can you send me a PM ?

texium wrote:

i can send it to you privately.

If possible. I also want to get these weights.

texium wrote:

Brendan wrote:

texium wrote:

And this game, is just beautiful.

Could you point me to where I can find this interesting Leela Tal 10k steps net?

i can send it to you privately.

Please do. Would be much appreciated.

here's an EAS update with much more games and a larger pool. its more aggressive than originally thought, and elo stat has it around 27-2800 CCRL.
https://www.mediafire.com/file/52o7jgkmkq3yp4y/leela-tal-10k.pb.gz/file

texium wrote:

here's an EAS update with much more games and a larger pool. its more aggressive than originally thought, and elo stat has it around 27-2800 CCRL.
https://www.mediafire.com/file/52o7jgkmkq3yp4y/leela-tal-10k.pb.gz/file

How many games were played?

150 wins and idk how many draws

A logo for Leela TAL (thks, Texium)

I see Leela-10 is a NN, can it be used in the CPU version of Leela?

yes, this is possible. However, a NN that was only trained with 10,000 training games is very weak. I doubt the origin of this NN.

Dio wrote:

yes, this is possible. However, a NN that was only trained with 16,000 training games is very weak. I doubt the origin of this NN.

Actually it's plenty strong considering it was only trained on human games to 10k steps. Badgyal has a hybrid of human and stockfish and was only 3400.

Dio wrote:

yes, this is possible. However, a NN that was only trained with 16,000 training games is very weak. I doubt the origin of this NN.

Well, that's one those keys for a high EAS

But at least with the CPU version one can play with a high concurrency.

[pgn][Event "elo"]
[Site "Lucas Chess R 2.14l"]
[Date "2024.10.15"]
[Round "?"]
[White "10k"]
[Black "Houdini 1.5a"]
[Result "1-0"]
[ECO "A21"]
[WhiteElo "2700"]
[BlackElo "3093"]
[PlyCount "195"]
[EventDate "2024.??.??"]

1. c4 e5 2. Nc3 d6 3. d4 exd4 4. Qxd4 Nc6 5. Qd2 a5 6. g3 Be6 7. Bg2 Bxc4 8. Nf3 Nf6 9. O-O Be7 10. b3 Ba6 11. Bb2 O-O 12. Nh4 Qd7 13. e4 Bxf1 14. Rxf1 g6 15. f4 Nb4 16. f5 c6 17. Bh3 Bd8 18. Nd1 Qe8 19. Nf2 Nxe4 20. Qh6 Nf6 21. Nf3 Qe2 {[#]} 22. fxg6 hxg6 {[#]} 23. Ng4 Qxb2 24. Ng5 Nbd5 {[#]} 25. Rxf6 Qc1+ 26. Bf1 Qxg5 27. Qxg5 Nxf6 28. Nxf6+ Kg7 29. Nh5+ Kh7 30. Qg4 f5 31. Qd1 gxh5 32. Qxh5+ Kg7 33. Bh3 Bb6+ 34. Kg2 Rf6 35. Bxf5 Rh8 36. Qg4+ Kf7 37. h4 d5 38. h5 Be3 39. Bd3 Rd8 40. Kh3 b5 41. a4 bxa4 42. Qxa4 Bb6 43. b4 axb4 44. Bg6+ Ke7 45. Qxb4+ c5 46. Qa3 Rh8 47. Qc3 d4 48. Qd3 Kd6 49. Qe4 Bd8 {[#]} 50. Qa8 Rff8 51. Qa6+ Kc7 52. Qe6 Rf6 53. Qe5+ Kc6 54. Be4+ Kb5 55. g4 Rfh6 56. Bd3+ Kc6 57. Be4+ Kb5 58. Qd5 Ba5 59. Kg2 Bc3 60. Bd3+ Kb6 61. Bg6 Bb4 62. Qd7 Bc3 63. Bd3 Bb4 64. Bc4 R8h7 65. Qd8+ Kc6 66. Qb8 Rh8 67. Bb5+ Kd5 68. Qb7+ Ke5 69. Qg7+ Ke4 70. Bc4 Kf4 71. Qf7+ Kxg4 72. Be6+ Rxe6 73. Qxe6+ Kxh5 74. Qe5+ Kg4 75. Qxh8 Kf4 76. Kf2 Ke4 77. Qg8 Kd3 78. Qb3+ Ke4 79. Ke2 Ke5 80. Kd3 Kd6 81. Kc4 Ke6 82. Qg3 Bd2 83. Kxc5 Be3 84. Kc6 Kf5 85. Kd5 Bf4 86. Qh3+ Kf6 87. Kxd4 Bg5 88. Ke4 Bd2 89. Qf5+ Ke7 90. Qe5+ Kd7 91. Qd5+ Kc7 92. Qxd2 Kc6 93. Qb4 Kd7 94. Kd5 Kc7 95. Qb5 Kd8 96. Qb7 Ke8 97. Ke6 Kd8 98. Qd7# 1-0


[/pgn]

Here the engine beats houdini 1.5 with great sacrifices

texium wrote:

here's an EAS update with much more games and a larger pool. its more aggressive than originally thought, and elo stat has it around 27-2800 CCRL.
https://www.mediafire.com/file/52o7jgkmkq3yp4y/leela-tal-10k.pb.gz/file

Is there a special version of Leela needed for this? I can't get it to work.