From c1575003e08abfba9cfe3db9c031b141cb0efa8c Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Sat, 21 Feb 2026 21:02:40 +0800
Subject: [PATCH] smaller init

---
 antipasto/peft_utils/antipasto_adapter.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/antipasto/peft_utils/antipasto_adapter.py b/antipasto/peft_utils/antipasto_adapter.py
index ea805ab..ab04094 100644
--- a/antipasto/peft_utils/antipasto_adapter.py
+++ b/antipasto/peft_utils/antipasto_adapter.py
@@ -279,7 +279,7 @@ class AntiPaSTOLayer(BaseTunerLayer):
             """With contrastive steering coeff=+1 and coeff=-1 produce identical outputs initially, so gradients are zero. Small random init is important for learning as it breaks symmetry."""
             x = torch.zeros(*args, **kwargs)
             # Option B: Draw from skew-symmetric distribution directly
-            nn.init.trunc_normal_(x, std=0.003)
+            nn.init.trunc_normal_(x, std=1e-4)
             x = x - x.T
             return x