Dev Builds » 20231230-1008

Use this dev build

NCM plays each Stockfish dev build 20,000 times against Stockfish 15. This yields an approximate Elo difference and establishes confidence in the strength of the dev builds.

Summary

Host Duration Avg Base NPS Games WLD Standard Elo Ptnml(0-2) Gamepair Elo
ncm-dbt-01 06:55:10 582520 4000 1414 662 1924 +66.1 ± 4.98 1 153 947 891 8 +135.96 ± 11.01
ncm-dbt-02 06:54:59 585065 4004 1450 633 1921 +71.9 ± 5.08 0 155 888 946 13 +147.85 ± 11.41
ncm-dbt-03 06:54:09 585159 4000 1382 658 1960 +63.59 ± 4.99 2 159 958 875 6 +130.94 ± 10.95
ncm-dbt-04 06:54:11 570820 4000 1395 659 1946 +64.66 ± 5.15 6 166 925 892 11 +133.14 ± 11.17
ncm-dbt-05 06:55:26 582483 3996 1396 669 1931 +63.92 ± 5.08 0 173 936 878 11 +130.29 ± 11.1
20000 7037 3281 9682 +66.03 ± 2.26 9 806 4654 4482 49 +135.6 ± 4.98

Test Detail

ID Host Base NPS Games WLD Standard Elo Ptnml(0-2) Gamepair Elo CLI PGN
407598 ncm-dbt-02 587028 4 2 1 1 +88.51 ± 137.63 0 0 1 1 0 +190.5 ± 634.21
407597 ncm-dbt-01 582402 500 192 92 216 +70.44 ± 13.91 0 18 114 118 0 +147.19 ± 31.9
407596 ncm-dbt-05 579166 496 166 90 240 +53.66 ± 14.4 0 22 132 90 4 +103.85 ± 29.33
407595 ncm-dbt-03 585295 500 182 85 233 +68.27 ± 14.4 0 22 109 119 0 +142.26 ± 32.74
407594 ncm-dbt-04 571391 500 179 95 226 +58.93 ± 15.6 2 23 119 101 5 +116.78 ± 31.26
407593 ncm-dbt-02 584076 500 175 78 247 +68.27 ± 14.54 0 22 110 117 1 +140.62 ± 32.58
407592 ncm-dbt-05 582527 500 177 79 244 +68.99 ± 13.86 0 17 119 113 1 +142.26 ± 31.1
407591 ncm-dbt-03 585632 500 172 78 250 +66.1 ± 14.03 0 19 119 111 1 +135.76 ± 31.16
407590 ncm-dbt-01 584411 500 182 79 239 +72.61 ± 14.27 0 17 116 114 3 +147.19 ± 31.57
407589 ncm-dbt-04 571833 500 174 81 245 +65.38 ± 14.29 0 20 119 109 2 +132.54 ± 31.18
407588 ncm-dbt-02 584327 500 180 83 237 +68.27 ± 14.12 0 19 116 114 1 +140.62 ± 31.61
407587 ncm-dbt-03 584664 500 175 77 248 +68.99 ± 13.41 0 15 122 113 0 +143.89 ± 30.58
407586 ncm-dbt-05 582110 500 180 87 233 +65.38 ± 14.56 0 23 112 114 1 +134.15 ± 32.28
407585 ncm-dbt-04 571592 500 181 99 220 +57.5 ± 14.08 1 19 128 101 1 +118.33 ± 29.86
407584 ncm-dbt-01 580406 500 165 82 253 +58.21 ± 14.25 0 24 119 107 0 +119.89 ± 31.25
407583 ncm-dbt-02 586308 500 174 74 252 +70.43 ± 14.05 0 18 115 116 1 +145.54 ± 31.75
407582 ncm-dbt-03 585169 500 167 85 248 +57.5 ± 14.36 1 20 127 100 2 +116.78 ± 30.03
407581 ncm-dbt-05 586985 500 168 82 250 +60.36 ± 14.89 0 27 111 111 1 +123.02 ± 32.44
407580 ncm-dbt-04 569909 500 172 77 251 +66.82 ± 14.48 1 19 115 114 1 +138.99 ± 31.79
407579 ncm-dbt-01 582861 500 178 71 251 +75.52 ± 13.49 0 14 115 121 0 +158.93 ± 31.63
407578 ncm-dbt-02 583405 500 185 78 237 +75.52 ± 14.23 0 17 111 120 2 +155.54 ± 32.36
407577 ncm-dbt-03 585211 500 171 81 248 +63.23 ± 13.63 0 18 124 108 0 +130.94 ± 30.38
407576 ncm-dbt-05 584327 500 180 83 237 +68.27 ± 14.12 0 19 116 114 1 +140.62 ± 31.61
407575 ncm-dbt-01 583363 500 177 94 229 +58.21 ± 14.79 1 23 120 104 2 +118.33 ± 31.1
407574 ncm-dbt-02 587240 500 189 93 218 +67.54 ± 14.37 0 21 113 115 1 +138.99 ± 32.11
407573 ncm-dbt-04 569350 500 162 79 259 +58.21 ± 14.39 1 22 120 107 0 +121.46 ± 31.09
407572 ncm-dbt-03 583447 500 177 83 240 +66.1 ± 14.59 1 20 114 114 1 +137.37 ± 31.96
407571 ncm-dbt-05 582945 500 179 89 232 +63.23 ± 14.61 0 25 110 115 0 +130.94 ± 32.59
407570 ncm-dbt-01 584034 500 175 79 246 +67.54 ± 14.23 0 19 118 111 2 +137.37 ± 31.31
407569 ncm-dbt-04 570228 500 183 77 240 +74.79 ± 14.34 0 18 110 120 2 +153.86 ± 32.54
407568 ncm-dbt-02 585337 500 178 64 258 +80.63 ± 14.38 0 17 104 127 2 +167.53 ± 33.52
407567 ncm-dbt-03 585084 500 178 89 233 +62.51 ± 14.17 0 21 120 108 1 +127.76 ± 31.05
407566 ncm-dbt-01 577479 500 170 79 251 +63.95 ± 13.94 0 20 119 111 0 +132.54 ± 31.18
407565 ncm-dbt-05 581485 500 174 82 244 +64.66 ± 13.24 0 15 128 107 0 +134.15 ± 29.68
407564 ncm-dbt-04 570629 500 178 78 244 +70.44 ± 14.48 1 19 109 121 0 +148.85 ± 32.73
407563 ncm-dbt-02 585295 500 190 87 223 +72.61 ± 14.55 0 20 109 119 2 +148.85 ± 32.73
407562 ncm-dbt-03 586774 500 160 80 260 +56.07 ± 14.29 0 24 123 102 1 +113.68 ± 30.67
407561 ncm-dbt-04 571632 500 166 73 261 +65.38 ± 14.83 0 26 105 119 0 +135.76 ± 33.36
407560 ncm-dbt-01 585211 500 175 86 239 +62.51 ± 13.74 0 18 126 105 1 +127.76 ± 30.09
407559 ncm-dbt-05 580323 500 172 77 251 +66.82 ± 15.16 0 25 108 114 3 +134.15 ± 32.9
407558 ncm-dbt-02 582569 500 177 75 248 +71.88 ± 14.8 0 21 109 117 3 +145.54 ± 32.74

Commit

Commit ID f12035c88c58a5fd568d26cde9868f73a8d7b839
Author Linmiao Xu
Date 2023-12-30 10:08:03 UTC
Update default net to nn-b1e55edbea57.nnue Created by retraining the master big net `nn-0000000000a0.nnue` on the same dataset with the ranger21 optimizer and more WDL skipping at training time. More WDL skipping is meant to increase lambda accuracy and train on fewer misevaluated positions where position scores are unlikely to correlate with game outcomes. Inspired by: - repeated reports in discord #events-discuss about SF misplaying due to wrong endgame evals, possibly due to Leela's endgame weaknesses reflected in training data - an attempt to reduce the skewed dataset piece count distribution where there are much more positions with less than 16 pieces, since the target piece count distribution in the trainer is symmetric around 16 The faster convergence seen with ranger21 is meant to: - prune experiment ideas more quickly since fewer epochs are needed to reach elo maxima - research faster potential trainings by shortening each run ```yaml experiment-name: 2560-S7-Re-514G-ranger21-more-wdl-skip training-dataset: /data/S6-514G.binpack early-fen-skipping: 28 start-from-engine-test-net: True nnue-pytorch-branch: linrock/nnue-pytorch/r21-more-wdl-skip num-epochs: 1200 lr: 4.375e-4 gamma: 0.995 start-lambda: 1.0 end-lambda: 0.7 ``` Experiment yaml configs converted to easy_train.sh commands with: https://github.com/linrock/nnue-tools/blob/4339954/yaml_easy_train.py Implementations based off of Sopel's NNUE training & experimentation log: https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY - Experiment 336 - ranger21 https://github.com/Sopel97/nnue-pytorch/tree/experiment_336 - Experiment 351 - more WDL skipping The version of the ranger21 optimizer used is: https://github.com/lessw2020/Ranger21/blob/b507df6/ranger21/ranger21.py The dataset is the exact same as in: https://github.com/official-stockfish/Stockfish/pull/4782 Local elo at 25k nodes per move: nn-epoch619.nnue : 6.2 +/- 4.2 Passed STC: https://tests.stockfishchess.org/tests/view/658a029779aa8af82b94fbe6 LLR: 2.93 (-2.94,2.94) <0.00,2.00> Total: 46528 W: 11985 L: 11650 D: 22893 Ptnml(0-2): 154, 5489, 11688, 5734, 199 Passed LTC: https://tests.stockfishchess.org/tests/view/658a448979aa8af82b95010f LLR: 2.94 (-2.94,2.94) <0.50,2.50> Total: 265326 W: 66378 L: 65574 D: 133374 Ptnml(0-2): 153, 30175, 71254, 30877, 204 This was additionally tested with the latest DualNNUE and passed SPRTs: Passed STC vs. https://github.com/official-stockfish/Stockfish/pull/4919 https://tests.stockfishchess.org/tests/view/658bcd5c79aa8af82b951846 LLR: 2.93 (-2.94,2.94) <0.00,2.00> Total: 296128 W: 76273 L: 75554 D: 144301 Ptnml(0-2): 1223, 35768, 73617, 35979, 1477 Passed LTC vs. https://github.com/official-stockfish/Stockfish/pull/4919 https://tests.stockfishchess.org/tests/view/658c988d79aa8af82b95240f LLR: 2.95 (-2.94,2.94) <0.50,2.50> Total: 75618 W: 19085 L: 18680 D: 37853 Ptnml(0-2): 45, 8420, 20497, 8779, 68 closes https://github.com/official-stockfish/Stockfish/pull/4942 Bench: 1304666
Copyright 2011–2025 Next Chess Move LLC