unembedding LR 0.004 to 0.006

2026-03-08 04:48:14 +00:00
parent 772dada6cc
commit aa8f408f39
1 changed files with 1 additions and 1 deletions
@@ -436,7 +436,7 @@ 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)
-UNEMBEDDING_LR = 0.004  # learning rate for lm_head (Adam)
+UNEMBEDDING_LR = 0.006  # 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)
 WEIGHT_DECAY = 0.2      # cautious weight decay for Muon