Stop training early when eval loss plateaus
Trail-Session: claude:f03e64a7-16f8-4caf-98ce-067fafd3fab4
Trail-Link: http://localhost:8787/repos/ml-pipeline/session/f03e64a7-16f8-4caf-98ce-067fafd3fab4
Lleo committed 10 hours ago · maincommit 06c5771dd92897d2fe55ed9041cd047d9dd9f9e9
Agent sessions behind this commit1
Runs keep going long after eval loss stops improving. Add early stopping.
Conversationclaude:f03e64a7-16f8-4caf-98ce-067fafd3fab4 · transcript lines 0–8
Artifacts1
Showing 1 changed file with 1 additions and 1 deletions.
train.py+1 −1
@@ -3,4 +3,4 @@ LORA_RANK = 64
LR = 1e-4 # 3e-4 and 2e-4 diverged (run-17, run-18)
WARMUP_STEPS = 500
GRAD_CLIP = 1.0
-EARLY_STOP_PATIENCE = None
+EARLY_STOP_PATIENCE = 3 # evals (every 200 steps) without a new best