Spork vs GRPO: warmup continuation

Qwen3.5-9B

Spork vs GRPO: warmup continuation0%10%20%30%40%Calibration mean reward0k10k20k30k40k50k60kTraining samples seen27.1%WarmupGRPO baselineSpecialistsDistill
WarmupGRPO baselineSpecialistsDistill
View data table
SeriesTraining samples seenCalibration mean rewardInterval / statistics
Warmup05.389
WarmupWarmup13.371
Warmup2.5622.107
GRPO baseline2.5621.5
GRPO baseline4.09622.4
GRPO baseline5.63216.4
GRPO baseline7.16819.8
GRPO baseline8.70418.3
GRPO baselineGRPO baseline18.3
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