Spork: Lichess held-out comparison

Qwen3.5-9B · 400 held-out puzzles (rating 800–1199) × 4 samples · best calibration checkpoint

Spork: Lichess held-out comparison0102030405060ScoreSolved (pass@1)Mean reward3.524.8423.8844.697.3734.230.8651.93
Base Qwen3.5-9BGRPOMOPD N=4, base-initSpork
View data table
SeriesCategoryScoreInterval / statistics
Base Qwen3.5-9BSolved (pass@1)3.5
GRPOSolved (pass@1)24.84
MOPD N=4, base-initSolved (pass@1)23.88
SporkSolved (pass@1)44.69
Base Qwen3.5-9BMean reward7.37
GRPOMean reward34.2
MOPD N=4, base-initMean reward30.86
SporkMean reward51.93