Spork vs GRPO full training run

Qwen3.5-9B · 400 calibration puzzles (rating 808–2776) · same 8,000 training puzzles

Spork vs GRPO full training run010203040Calibration mean reward0k10k20k30k40k50k60k70kTraining samples seen27.1GRPO baselineWarmupSpecialistsDistill
GRPO baselineWarmupSpecialistsDistill
View data table
SeriesTraining samples seenCalibration mean rewardInterval / statistics
GRPO baseline05.389
GRPO baseline8.1926.267
GRPO baseline16.38410.014
GRPO baseline24.57610.684
GRPO baseline32.76815.231
GRPO baseline40.9618.104
GRPO baseline49.15222.576
GRPO baselineGRPO baseline23.361
GRPO baseline65.53623.349
GRPO baseline6821.5
GRPO baseline7020.9
Warmup05.389
WarmupWarmup13.371
Specialists1.2813.371
Specialists2.5622.107
Specialists4.09621.718
Specialists5.63219.943
Specialists7.16820.759
Specialists8.70422.724
Specialists10.2423.283
Specialists11.77623.588
Specialists13.31223.765
Specialists14.84824.805
Specialists16.38424.563
Specialists17.9225.434
Specialists26.11227.277
Specialists34.30426.538
SpecialistsSpecialists27.209
Specialists43.5226.57
Specialists48.12826.198
Specialists52.73627.054
Distill06
Distill1.53622.312
Distill3.07227.799
Distill4.60830.998
Distill6.14431.039
Distill7.6829.831
Distill9.21630.356
Distill10.75230.766
Distill12.28831.601
DistillDistill31.894