-
9fc61ca82b
configure for group < 10 trace grpo training
master
yuyr
2025-07-01 09:50:29 +08:00
-
4295f30f9a
update direct stepwise train, pass trainning with mock one step trace duplicated 52 times; reward OK, curve improved as the step grows.
yuyr
2025-06-30 15:06:38 +00:00
-
5e632c53ac
temp save; temp hold on for rejection sample CoT; try direct RL with raw stepwise data first
yuyr
2025-06-30 10:29:00 +00:00
-
7f4fc8b05b
use --system for system prompt
yuyr
2025-06-26 22:46:09 +00:00
-
ee08da12c0
pass sample test web with custom orm
yuyr
2025-06-26 18:02:13 +00:00
-
8f168ecbef
test swift for qwen3 math
yuyr
2025-06-25 17:00:47 +08:00