a3-rl-DCAgent_mix_h2_language_proportional-65-8B/training_logs/report.md

RL Training Report — a3-rl-DCAgent_mix_h2_language_proportional

Chain 570909-570912 (winner, reached max_steps=80). EMA-best (alpha=1/3) over global_step<=80.

Selected checkpoint: global_step_65 EMA=0.4924 reward=0.6816 pass@8=0.734375

step	reward	EMA	pass@8	grad_norm	entropy
10	0.3086	0.2947	0.46875	0.012359619140625	NA
15	0.4180	0.3842	0.609375	0.0121917724609375	NA
20	0.4609	0.3983	0.609375	0.008453369140625	NA
25	0.5098	0.4421	0.65625	0.00807952880859375	NA
30	0.4805	0.4224	0.578125	0.00757598876953125	NA
35	0.4238	0.4208	0.546875	0.00506591796875	NA
40	0.3301	0.4021	0.421875	0.00420379638671875	NA
45	0.3809	0.3868	0.5625	0.00640869140625	NA
50	0.3867	0.3979	0.5	0.00432586669921875	NA
55	0.3730	0.3763	0.484375	0.00620269775390625	NA
60	0.4121	0.4032	0.578125	0.00632476806640625	NA
65	0.6816	0.4924	0.734375	0.00507354736328125	NA
70	0.5273	0.4775	0.640625	0.0052642822265625	NA
75	0.3555	0.4352	0.578125	0.006622314453125	NA
80	0.4746	0.4275	0.65625	0.00689697265625	NA

step 80: reward=0.4746 pass@8=0.65625