[Tunix] Default degenerate_group_masking to false.

lc5211 · The tunix Authors · commit ef192de79f18 · 2026-06-03T10:36:47.000-07:00
PiperOrigin-RevId: 926128256
diff --git a/tunix/rl/agentic/agentic_grpo_learner.py b/tunix/rl/agentic/agentic_grpo_learner.py
@@ -107,7 +107,7 @@ class GRPOConfig(agentic_rl_learner.AgenticRLConfig):
   epsilon_high: float | None = None  # 0.28 from DAPO.
   off_policy_steps: int = 0
   degenerate_group_masking: bool = (
-      True  # Whether to mask out degenerate groups with all-0 advantages.
+      False  # Whether to mask out degenerate groups with all-0 advantages.
   )
   use_rollout_logps: bool = True
   # Truncated importance-sampling (TIS) correction for the residual mismatch

Original file line number	Diff line number	Diff line change
`@@ -107,7 +107,7 @@ class GRPOConfig(agentic_rl_learner.AgenticRLConfig):`
`107`	`107`	`epsilon_high: float \| None = None # 0.28 from DAPO.`
`108`	`108`	`off_policy_steps: int = 0`
`109`	`109`	`degenerate_group_masking: bool = (`
`110`		`- True # Whether to mask out degenerate groups with all-0 advantages.`
	`110`	`+ False # Whether to mask out degenerate groups with all-0 advantages.`
`111`	`111`	`)`
`112`	`112`	`use_rollout_logps: bool = True`
`113`	`113`	`# Truncated importance-sampling (TIS) correction for the residual mismatch`