From 7d047e42f42cafb0cb2609d53ce0b2bd0590afce Mon Sep 17 00:00:00 2001
From: autoresearch <autoresearch@local>
Date: Sun, 8 Mar 2026 04:19:15 +0000
Subject: [PATCH] embedding LR 0.6 to 0.8

---
 train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train.py b/train.py
index fdef2cb..c14b3e0 100644
--- a/train.py
+++ b/train.py
@@ -435,7 +435,7 @@ WINDOW_PATTERN = "SSSSL" # sliding window pattern: L=full, S=half context
 
 # Optimization
 TOTAL_BATCH_SIZE = 2**18 # ~262K tokens per optimizer step
-EMBEDDING_LR = 0.6      # learning rate for token embeddings (Adam)
+EMBEDDING_LR = 0.8      # learning rate for token embeddings (Adam)
 UNEMBEDDING_LR = 0.004  # learning rate for lm_head (Adam)
 MATRIX_LR = 0.04        # learning rate for matrix parameters (Muon)
 SCALAR_LR = 0.5         # learning rate for per-layer scalars (Adam)