Spork vs teachers: Lichess partitions

Qwen3.5-9B · eval partition split

Spork vs teachers: Lichess partitions01020304050607080Held-out mean rewardKing attack / mateForcing tacticsLine / materialtacticsQuiet conversion6.430.931.120.541.341.39.937.915.236.549.358.06.433.023.230.941.746.09.539.134.139.254.763.8
Base Qwen3.5-9BGRPOBase-init teachersBase-init MOPDSpork teachersSpork
View data table
SeriesCategoryHeld-out mean rewardInterval / statistics
Base Qwen3.5-9BKing attack / mate6.4
GRPOKing attack / mate30.9
Base-init teachersKing attack / mate31.1
Base-init MOPDKing attack / mate20.5
Spork teachersKing attack / mate41.3
SporkKing attack / mate41.3
Base Qwen3.5-9BForcing tactics9.9
GRPOForcing tactics37.9
Base-init teachersForcing tactics15.2
Base-init MOPDForcing tactics36.5
Spork teachersForcing tactics49.3
SporkForcing tactics58
Base Qwen3.5-9BLine / material tactics6.4
GRPOLine / material tactics33
Base-init teachersLine / material tactics23.2
Base-init MOPDLine / material tactics30.9
Spork teachersLine / material tactics41.7
SporkLine / material tactics46
Base Qwen3.5-9BQuiet conversion9.5
GRPOQuiet conversion39.1
Base-init teachersQuiet conversion34.1
Base-init MOPDQuiet conversion39.2
Spork teachersQuiet conversion54.7
SporkQuiet conversion63.8