Upload train.py with huggingface_hub

train.py

@@ -131,6 +131,20 @@ def train(a):
     recs = [r for r in recs if r["src"] not in gold_ids]   # never train on gold
     print("training %d records | gold %d (disjoint)" % (len(recs), len(gold)), flush=True)
 
+    # --init is NOT --resume. Resuming continues an interrupted run on its own
+    # schedule; init starts a NEW run from trained weights, with a fresh optimizer
+    # and a fresh learning-rate curve. Use it to add a question, or to retrain one
+    # question on a rebalanced set without starting from the base model.
+    #
+    # This has to happen BEFORE the baseline measurement. Loading the weights
+    # after it prints a "before" score for the untrained model -- which is not
+    # the number you are comparing against, and the run looks like a bigger win
+    # than it was.
+    if a.init:
+        st = torch.load(a.init, map_location=dev, weights_only=False)
+        m.load_state_dict(st["model"] if "model" in st else st)
+        print("  initialised from %s (fresh optimizer and schedule)" % a.init, flush=True)
+
     print("BEFORE:", json.dumps(evaluate(m, gold, tok, cfg, dev)), flush=True)
     if a.eval_only:
         return
@@ -173,15 +187,6 @@ def train(a):
     # two running moments per parameter, and OneCycleLR has to know where on its
     # curve it stopped. Without both it is a warm start with a fresh warmup cycle,
     # not a continuation -- the learning rate jumps back up and shakes the weights.
-    # --init is NOT --resume. Resuming continues an interrupted run on its own
-    # schedule; init starts a NEW run from trained weights, with a fresh optimizer
-    # and a fresh learning-rate curve. Use it to add a question, or to retrain one
-    # question on a rebalanced set without starting from the base model.
-    if a.init:
-        st = torch.load(a.init, map_location=dev, weights_only=False)
-        m.load_state_dict(st["model"] if "model" in st else st)
-        print("  initialised from %s (fresh optimizer and schedule)" % a.init, flush=True)
-
     src = a.resume or os.environ.get("LAYA_RESUME")
     if src and os.path.exists(src):
         st = torch.load(src, map_location=dev, weights_only=False)