Commit 0e47e9b
committed
Update default Fischer Network to [408F1C41]
This network is applying what we've learned about WDL, to the Fischer networks. The gains here are strictly over the now previous best Fischer network, s46.FRC.8x32.e390.e60.x768.nn [AD09A846]
Summary of all training to get to this point...
Step 1. We trained AD09A846, using the s46. Trained for 390 epochs (1024x1024x128 steps each), at LR=0.001.
Step 2. Then 60 epochs at LR=0.000025, a 40x drop.
Step 3. Convert AD09A846 back into the trainer format. Losing pre-quant values, factorized weights, and ADAM state.
Step 4. We trained 408F1C41 using an additional 55 epochs of (100% WDL, 0% Eval) on the s46 data.
The s46 data, is 100m N=40k FRC, trained, using the 2x512 network B5E1899E. It contains 7,509,194,191 positions.
Standard Chess ( Just to see )...
Elo | 29.91 +- 5.67 (95%)
Conf | 60.0+0.60s Threads=1 Hash=64MB
Games | N: 3424 W: 926 L: 632 D: 1866
Penta | [0, 268, 891, 544, 9]
http://chess.grantnet.us/test/37224/
With 3moves_FRC.epd ( Very drawish )...
Elo | 16.52 +- 1.93 (95%)
Conf | 60.0+0.60s Threads=1 Hash=64MB
Games | N: 16860 W: 1958 L: 1157 D: 13745
Penta | [7, 712, 6222, 1451, 38]
http://chess.grantnet.us/test/37236/
Bench 24921871 parent b1b13a5 commit 0e47e9b
1 file changed
+1
-1
lines changed| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
23 | 23 | | |
24 | 24 | | |
25 | 25 | | |
26 | | - | |
| 26 | + | |
27 | 27 | | |
28 | 28 | | |
29 | 29 | | |
| |||
0 commit comments