diff --git a/app/web/main.py b/app/web/main.py
index c6e3c33..391b32b 100644
--- a/app/web/main.py
+++ b/app/web/main.py
@@ -914,6 +914,18 @@ for _m in DATASET_META.values():
     _GEN_TO_META.setdefault(_m["path"].rsplit(".", 1)[-1], _m)
 
 
+# Kwargs the flow injects / we supply explicitly — never part of the
+# dataset's semantic identity, so strip them before DATASET_META matching
+# and before regenerating labels.
+_TRANSIENT_GEN_KWARGS = {"n_samples", "random_state"}
+
+
+def _clean_gen_kwargs(gk: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
+    if gk is None:
+        return None
+    return {k: v for k, v in gk.items() if k not in _TRANSIENT_GEN_KWARGS}
+
+
 def _lookup_dataset_meta(
     generator_short: str, generator_kwargs: Optional[Dict[str, Any]]
 ) -> Optional[Dict[str, Any]]:
@@ -926,9 +938,10 @@ def _lookup_dataset_meta(
     ]
     if not candidates:
         return None
-    if generator_kwargs is not None:
+    gk = _clean_gen_kwargs(generator_kwargs)
+    if gk is not None:
         for m in candidates:
-            if m["kwargs"] == generator_kwargs:
+            if m["kwargs"] == gk:
                 return m
     return candidates[0]
 
@@ -963,7 +976,7 @@ def _enrich_with_labels(d: Dict[str, Any]) -> Dict[str, Any]:
     dm = _lookup_dataset_meta(gen_short, gk)
     if not dm:
         return d
-    kwargs_to_use = gk if gk is not None else dm["kwargs"]
+    kwargs_to_use = _clean_gen_kwargs(gk) if gk is not None else dm["kwargs"]
     try:
         mod_path, cls_name = dm["path"].rsplit(".", 1)
         fn = getattr(importlib.import_module(mod_path), cls_name)