SolsticeBench

Finished
TestTCEngineAuthorResult
net-l80ab_ml_sb640 SPRT diff
DIAGNOSTIC: ml net at equal nodes, 400k/move (8x #241): does the per-node edge hold at deeper search?
N=400000SolsticeDragonLLR: 2.94 (-2.94, 2.94) [0.00, 5.00]
Games: 1364 W: 423 L: 300 D: 641
Ptnml(0-2): 9, 114, 317, 229, 13
r50-damp SPRT diff
Static eval x (200 - halfmove)/200 (SF rule50 damping)
4.0+0.04SolsticeDragonLLR: -1.48 (-2.94, 2.94) [0.00, 5.00]
Games: 5944 W: 1464 L: 1496 D: 2984
Ptnml(0-2): 33, 717, 1508, 677, 37
optimism2 SPRT diff
Optimism (SF): opt_k*s/(|s|+opt_d) from previous iteration, scaled by (20+material)/240
4.0+0.04SolsticeDragonLLR: -1.84 (-2.94, 2.94) [0.00, 5.00]
Games: 6976 W: 1695 L: 1737 D: 3544
Ptnml(0-2): 55, 844, 1718, 830, 41
cuckoo2 SPRT diff
search: cuckoo upcoming-repetition detection (alpha raised to draw when one reversible move repeats a position of the line)
4.0+0.04SolsticeDragonLLR: -2.44 (-2.94, 2.94) [0.00, 5.00]
Games: 12682 W: 3099 L: 3140 D: 6443
Ptnml(0-2): 65, 1517, 3212, 1488, 59
conthist-kind SPRT diff
search: continuation history split by [in check][noisy] of the move keying each row (Stockfish/Reckless)
4.0+0.04SolsticeDragonLLR: -2.49 (-2.94, 2.94) [0.00, 5.00]
Games: 13116 W: 3254 L: 3296 D: 6566
Ptnml(0-2): 79, 1583, 3268, 1557, 71
corr-pair SPRT diff
search: eval correction keyed by the last two moves (piece,to)x(piece,to), weight 100%
4.0+0.04SolsticeDragonLLR: -2.95 (-2.94, 2.94) [0.00, 5.00]
Games: 9638 W: 2380 L: 2452 D: 4806
Ptnml(0-2): 52, 1182, 2421, 1114, 50
net-l80ab_ml_sb640 SPRT diff
First multi-layer net (16x32, CNU6 quant, sparse L1), ~0.55x nps
4.0+0.04SolsticeDragonLLR: -3.00 (-2.94, 2.94) [0.00, 5.00]
Games: 3350 W: 821 L: 923 D: 1606
Ptnml(0-2): 45, 420, 817, 378, 15
net-l80ab_ml_sb640 SPRT diff
DIAGNOSTIC: ml net at equal nodes (50k/move): net quality without the 0.55x speed
N=50000SolsticeDragonLLR: 2.99 (-2.94, 2.94) [0.00, 5.00]
Games: 2054 W: 704 L: 556 D: 794
Ptnml(0-2): 34, 214, 414, 300, 65
l80ab_ml_kb16_h1024_l16x32_w025_sb640 Training
h1024Solsticechess-engine-test-top-levelbest val 0.4602 (ep 600)
16 epochs · 16h 35m
lmr-singular SPRT diff
search: LMR of the TT move's siblings += 0.25 ply + min(singular margin * 10/1024, 1) after a singular TT move
4.0+0.04SolsticeDragonLLR: -2.97 (-2.94, 2.94) [0.00, 5.00]
Games: 7856 W: 1916 L: 1994 D: 3946
Ptnml(0-2): 30, 977, 1996, 891, 34
lmr-extend SPRT diff
search: quiet LMR may go to -1 (null-window search one ply deeper)
4.0+0.04SolsticeDragonLLR: -2.97 (-2.94, 2.94) [0.00, 5.00]
Games: 1740 W: 379 L: 487 D: 874
Ptnml(0-2): 14, 268, 410, 168, 10
capthist-threat SPRT diff
search: threat-indexed capture history (Reckless noisy history: victim*13.9 + hist, history-dependent SEE split)
4.0+0.04SolsticeDragonLLR: 2.98 (-2.94, 2.94) [0.00, 5.00]
Games: 9744 W: 2544 L: 2395 D: 4805
Ptnml(0-2): 48, 1129, 2385, 1246, 64
evaldiff3 SPRT diff
search: eval-difference history, calibrated (lopsided offset 250 malus<=60 bonus<=250, x0.5; history stats match master)
4.0+0.04SolsticeDragonLLR: -1.27 (-2.94, 2.94) [0.00, 5.00]
Games: 1234 W: 296 L: 338 D: 600
Ptnml(0-2): 6, 174, 296, 138, 3
spsa229 SPRT diff
SPSA tune #229 values (retune for the kb16ob8 h1024 net)
4.0+0.04SolsticeDragonLLR: -0.38 (-2.94, 2.94) [0.00, 5.00]
Games: 3646 W: 929 L: 929 D: 1788
Ptnml(0-2): 23, 433, 904, 447, 16
evaldiff3 SPRT diff
search: eval-difference history centred on the measured bias (offset 200, clamp +-200, x8)
4.0+0.04SolsticeDragonLLR: -0.83 (-2.94, 2.94) [0.00, 5.00]
Games: 832 W: 185 L: 214 D: 433
Ptnml(0-2): 7, 122, 182, 103, 2
master SPSA diff
SPSA retune for the final Leela net (kb16ob8 h1024), master 69c7219, all 87 params, 1000x32 @2+0.02
2.0+0.02SolsticeDragonTuning 87 Parameters
1000/1000 Iterations
64008/64000 Games Played
solstice-69c7219-ccrl-blitz Gauntlet
8 opponents · 8moves_v3 (balanced)
60+0.6Solstice 69c7219chess-engine-test-top-level3702 ± 16 CCRL Blitz
462 games · 39m
net-l80ab_kb16ob8_h1024_w025_sb640 SPRT diff
final Leela net: kb16 ob8 h1024, 4.8B Leela positions, 640 SB; gen6 val 0.46140 vs 0.46585
4.0+0.04SolsticeDragonLLR: 3.08 (-2.94, 2.94) [0.00, 5.00]
Games: 1004 W: 336 L: 203 D: 465
Ptnml(0-2): 5, 71, 236, 166, 24
l80ab_kb16ob8_h1024_w025_sb640 Training
h1024Solsticechess-engine-test-top-levelbest val 0.4614 (ep 640)
16 epochs · 15h 16m
89B5780FD0A2CAE9 SPRT diff
smp-skip: SF9-style depth-skip pattern for helpers (8 threads)
5.0+0.05SolsticeDragonLLR: -2.96 (-2.94, 2.94) [0.00, 5.00]
Games: 6910 W: 1586 L: 1665 D: 3659
Ptnml(0-2): 16, 864, 1773, 787, 15
AD85825116ACCAFE SPRT diff
smp-vote: best-thread vote (8 threads)
5.0+0.05SolsticeDragonLLR: -2.94 (-2.94, 2.94) [0.00, 5.00]
Games: 6642 W: 1580 L: 1659 D: 3403
Ptnml(0-2): 12, 837, 1699, 764, 9
D68F24583B4B8EB8 SPRT diff
smp-hist: helpers keep histories between moves (8 threads)
5.0+0.05SolsticeDragonLLR: 2.95 (-2.94, 2.94) [0.00, 5.00]
Games: 7002 W: 1741 L: 1610 D: 3651
Ptnml(0-2): 17, 783, 1774, 906, 21
solstice-2793cc7-ccrl-blitz Gauntlet
8 opponents · 8moves_v3 (balanced)
60+0.6Solstice 2793cc7chess-engine-test-top-level3686 ± 14 CCRL Blitz
640 games · 54m
15B4E162BFEA9263 SPRT diff
net: l80_kb4_h512_w025_sb400 (pure Leela data: 1.05B positions from 398 lc0 test80 2025 tars (leela2bullet, score 800*atanh(best_q)), kb4 h512, 25% result, 400 SB. Leela val 0.59633 vs master 0.60304; gen6 val 0.46585 vs 0.46098; eval slope 0.93)
4.0+0.04SolsticeDragonLLR: 3.05 (-2.94, 2.94) [0.00, 5.00]
Games: 1182 W: 412 L: 272 D: 498
Ptnml(0-2): 11, 101, 242, 211, 26
l80_kb4_h512_w025_sb400 Training
h512Solsticechess-engine-test-top-levelbest val 0.4658 (ep 400)
10 epochs · 3h 52m
gen7 Datagen
gen7-bulletformat-g456_kb4_h512-5k
5.0k nodesSolsticechess-engine-test-top-level617M positions
7.8M games · 0m
57CB2A2E82EEC17A SPRT diff
multicut-sf: SF multicut (s_score >= beta -> return s_score); local +12.2 +-10.0
4.0+0.04SolsticeDragonLLR: -0.29 (-2.94, 2.94) [0.00, 5.00]
Games: 7376 W: 1734 L: 1716 D: 3926
Ptnml(0-2): 55, 886, 1791, 898, 58
4308671917A0510D SPRT diff
tm80: SF-style falling-eval TM factor (prev search + 3-iter drop), mean-time normalised. Local +14.0 +-10.4 (2.4k) from the big-gainer checkers.
4.0+0.04SolsticeDragonLLR: 3.07 (-2.94, 2.94) [0.00, 5.00]
Games: 8416 W: 2049 L: 1902 D: 4465
Ptnml(0-2): 53, 919, 2124, 1052, 60
5D53C15199CB4A09 SPRT diff
bundle7: hindsight + 50-move damping + qs futility + see-hist + ext budget + se tt-pv margin + qs quiet queen promotions
4.0+0.04SolsticeDragonLLR: -3.12 (-2.94, 2.94) [0.00, 5.00]
Games: 7704 W: 1789 L: 1878 D: 4037
Ptnml(0-2): 45, 1021, 1822, 906, 58
F185A617DD7DDDE8 SPRT diff
root-order (features #10): root moves by previous node counts, TT move first
4.0+0.04SolsticeDragonLLR: -1.97 (-2.94, 2.94) [0.00, 5.00]
Games: 2560 W: 585 L: 651 D: 1324
Ptnml(0-2): 26, 344, 592, 306, 12
813C3CAF47FDD770 SPRT diff
threat-order2 (features #11): quiet ordering +-6000 for escaping/entering lesser-piece threats
4.0+0.04SolsticeDragonLLR: -0.50 (-2.94, 2.94) [0.00, 5.00]
Games: 2048 W: 452 L: 463 D: 1133
Ptnml(0-2): 8, 269, 486, 248, 13
g456_kb4ob8_h512_w025_sb400 Training
h512Solsticechess-engine-test-top-levelbest val 0.4607 (ep 400)
10 epochs · 4h 27m
g456_kb4_h1024_w025_sb400 Training
h1024Solsticechess-engine-test-top-levelbest val 0.4594 (ep 400)
10 epochs · 7h 53m
g456_kb4_h512_w025_sb800 Training
h512Solsticechess-engine-test-top-levelbest val 0.4608 (ep 800)
20 epochs · 8h 22m
g456_kb4_h512_w025_sb400 Training
h512Solsticechess-engine-test-top-levelbest val 0.4610 (ep 400)
20 epochs · 4h 12m
g456_h512_w025_sb400 Training
h512Solsticechess-engine-test-top-levelbest val 0.4630 (ep 380)
20 epochs · 3h 30m
gen6 Datagen
gen6-6f5e610-g45_h512-5k
5.0k nodesSolsticechess-engine-test-top-level910M positions
7.8M games · 0m
g45_kb4_h512_w025_sb240 Training
h512Solsticechess-engine-test-top-levelbest val 0.4528 (ep 240)
12 epochs · 1h 48m
solstice-60bfc00-ccrl-blitz Gauntlet
8 opponents · 8moves_v3 (balanced)
60+0.6Solstice 60bfc00chess-engine-test-top-level3549 ± 24 CCRL Blitz
252 games · 23m
g45_h512_w025_sb120 Training
h512Solsticechess-engine-test-top-levelbest val 0.4531 (ep 120)
12 epochs · 39m
g45_ob8_w025_sb320 Training
h256Solsticechess-engine-test-top-levelbest val 0.4540 (ep 320)
16 epochs · 59m
gen4_ob8_cos60 Training
h256 λ0.75Solsticechess-engine-test-top-levelbest val 0.4685 (ep 42)
42 epochs · 2h 32m
g45_w025_sb120 Training
h256Solsticechess-engine-test-top-levelbest val 0.4544 (ep 120)
12 epochs · 19m
g4_w010_sb60 Training
h256Solsticechess-engine-test-top-levelbest val 0.4572 (ep 60)
12 epochs · 9m
g4_w075_sb60 Training
h256Solsticechess-engine-test-top-levelbest val 0.4564 (ep 60)
1 epochs · 9m
g4_w050_sb60 Training
h256Solsticechess-engine-test-top-levelbest val 0.4552 (ep 60)
1 epochs · 9m
g4_w025_sb60 Training
h256Solsticechess-engine-test-top-levelbest val 0.4560 (ep 60)
1 epochs · 9m
fly2 Training
gen4 (Solstice self-play)
λ0.75Solsticechess-engine-test-top-levelbest val 0.4762 (ep 48)
49 epochs · 0m
gen45_kb4_cos40 Training
gen4+gen5
h256 λ0.75Solsticexeon-gpubest val 0.4792 (ep 40)
40 epochs · 4h 21m
gen5_h256_cos60 Training
cache_gen5
h256 λ0.75Solsticexeon-gpubest val 0.4897 (ep 60)
60 epochs · 3h 27m
gen5b Datagen
gen5b-19341ce-gen4_h256_cos60-10k
10.0k nodesSolsticechess-engine-test-top-level43.0M positions
333k games · 2h 38m
solstice-130739e-ccrl-blitz Gauntlet
8 opponents · 8moves_v3 (balanced)
60+0.6Solstice 130739echess-engine-test-top-level3420 ± 16 CCRL Blitz
640 games · 55m
fly1 Training
gen4 (Solstice self-play)
λ0.75Solsticechess-engine-test-top-levelbest val 0.4851 (ep 26)
44 epochs · 0m
pilot_fly Training
gen4 (Solstice self-play)
λ0.75Solsticechess-engine-test-top-levelbest val 0.5998 (ep 2)
3 epochs · 0m
gen5 Datagen
gen5-19341ce-gen4_h256_cos60-10k
10.0k nodesSolsticechess-engine-test-top-level100M positions
777k games · 0m
gen5a Datagen
gen5a-19341ce-gen4_h256_cos60-10k
10.0k nodesSolsticechess-engine-test-top-level844k positions
6.5k games · 0m
gen34_h512_cos40 Training
cache_gen34
h512 λ0.75Solsticexeon-gpubest val 0.4675 (ep 40)
40 epochs · 7h 47m
gen4_h256_cos60 Training
cache_gen4
h256 λ0.75Solsticexeon-gpubest val 0.4673 (ep 59)
60 epochs · 3h 26m
gen4 Datagen
9c793e0-gen4-gen3_h256_cos60
5.0k nodesSolsticechess-engine-test-top-level118M positions
986k games · 3h 18m
gen4b Datagen
gen4b-topup-gen3_h256_cos60
5.0k nodesSolsticechess-engine-test-top-level36.7M positions
307k games · 44m
gen3_h256_cos60_qat Training
cache_gen3
h256 λ0.75Solsticexeon-gpubest val 0.4694 (ep 9)
10 epochs · 30m
gen3_h256_cos60 Training
cache_gen3
h256 λ0.75Solsticexeon-gpubest val 0.4692 (ep 60)
60 epochs · 2h 56m
solstice-596e584-ccrl-blitz Gauntlet
8 opponents · 8moves_v3 (balanced)
60+0.6Solstice 596e584xeon-gpu3240 ± 27 CCRL Blitz
425 games · 39m
gen3_h256_l075 Training
cache_gen3
h256 λ0.75Solsticesolsticeboxbest val 0.4703 (ep 12)
15 epochs · 6h 28m
gen3_l075 Training
cache_gen3
h128 λ0.75Solsticesolsticeboxbest val 0.4710 (ep 14)
15 epochs · 5h 04m
gen2_l075 Training
cache_gen2
λ0.75Solsticesolsticeboxbest val 0.4651 (ep 15)
15 epochs · 35m
gen1_l075 Training
cache_gen1
λ0.75Solsticesolsticeboxbest val 0.4421 (ep 15)
15 epochs · 1h 38m
gen3 Datagen
48c5d29-gen3-solstice.nnue
5.0k nodesSolsticesolstice-xeon80135M positions
1.2M games · 4h 23m
gen2 Datagen
8269a3d-gen2-canonchess.nnue
5.0k nodesSolsticesolsticebox44.9M positions
385k games · 6h 56m
gen1 Datagen
814dc8e-gen1-canonchess.nnue
5.0k nodesSolsticesolsticebox44.9M positions
387k games · 7h 10m