From 49e813f83bfbde80849d532be9e5b1d4e65e2c13 Mon Sep 17 00:00:00 2001
From: Django-Jiang <43953876+Django-Jiang@users.noreply.github.com>
Date: Fri, 6 Mar 2026 22:38:41 -0800
Subject: [PATCH] [fix] Fix numerical accuracy issue in dynamic sampling filter

Use float64 instead of float32 and compare std against 1e-6 epsilon
instead of exact 0.0 to avoid false negatives from floating-point
precision errors.

Made-with: Cursor
---
 slime/rollout/filter_hub/dynamic_sampling_filters.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/slime/rollout/filter_hub/dynamic_sampling_filters.py b/slime/rollout/filter_hub/dynamic_sampling_filters.py
index dc17422d68..9c579d0fda 100644
--- a/slime/rollout/filter_hub/dynamic_sampling_filters.py
+++ b/slime/rollout/filter_hub/dynamic_sampling_filters.py
@@ -8,7 +8,7 @@
 
 def check_reward_nonzero_std(args, samples: list[Sample], **kwargs):
     rewards = [sample.get_reward_value(args) for sample in samples]
-    keep = torch.tensor(rewards, dtype=torch.float).std() > 0.0
+    keep = torch.tensor(rewards, dtype=torch.float64).std() > 1e-6
     return DynamicFilterOutput(
         keep=keep,
         reason=None if keep else f"zero_std_{round(rewards[0], 1)}",