Restrict leaderboard to live TS-Bench datasets

src/utils.py

@@ -125,6 +125,25 @@ def _read_result_csv(path: str | Path) -> pd.DataFrame:
     return df
 
 
+def _dataset_key(value: object) -> str:
+    return str(value).split("/")[0].strip().lower()
+
+
+def _load_live_dataset_keys(ds_properties: str | Path | None) -> set[str]:
+    if ds_properties is None:
+        return set()
+    path = Path(ds_properties)
+    if not path.exists():
+        return set()
+    try:
+        props = pd.read_csv(path)
+    except (OSError, pd.errors.ParserError, pd.errors.EmptyDataError):
+        return set()
+    if "dataset" not in props.columns:
+        return set()
+    return {_dataset_key(value) for value in props["dataset"].dropna()}
+
+
 GIFT_AGGREGATE_FILES = {
     "prediction_length": "results_by_prediction_length.csv",
     "domain": "results_by_domain.csv",
@@ -231,6 +250,11 @@ def prepare_results_df(
     df = load_all_results(root_dir)
     if df.empty:
         return df
+    live_dataset_keys = _load_live_dataset_keys(ds_properties)
+    if live_dataset_keys:
+        df = df[df["dataset"].map(_dataset_key).isin(live_dataset_keys)].copy()
+        if df.empty:
+            return pd.DataFrame(columns=[*RESULT_COLUMNS, "frequency", "term_length"])
     if Path(ds_properties).exists():
         df = enrich_with_properties(df, ds_properties)
     if "domain" in df.columns:
@@ -577,7 +601,7 @@ def lookup_dataset_domain(dataset: str, root_dir: str = "results") -> str:
     props_path = Path(root_dir) / "dataset_properties.csv"
     if props_path.exists():
         props = pd.read_csv(props_path)
-        match = props[props["dataset"].str.lower() == base.lower()]
+        match = props[props["dataset"].map(_dataset_key) == _dataset_key(base)]
         if not match.empty and "domain" in match.columns:
             return str(match.iloc[0]["domain"])