diff --git a/train.py b/train.py
index d1c6b9f..3ac6503 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 = 0.8      # learning rate for token embeddings (Adam)
+EMBEDDING_LR = 0.9      # learning rate for token embeddings (Adam)
 UNEMBEDDING_LR = 0.005  # 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)