fix(leaderboard): restore per-language drilldown on row click

hf-space/requirements.txt

@@ -1,4 +1,4 @@
-# The Space pulls the package from PyPI with the leaderboard extra so the
-# Gradio entrypoint and `commonlid leaderboard serve` CLI stay in sync. Pin a
-# concrete release here once we tag one (e.g. `commonlid[leaderboard]==0.1.0`).
-commonlid[leaderboard]
+# Install commonlid directly from GitHub main so the Space picks up changes
+# without a PyPI release cycle. Pin to a tag or commit SHA (e.g.
+# `...@v0.2.3` or `...@<sha>`) if you need to lock the Space to a known build.
+commonlid[leaderboard] @ git+https://github.com/commoncrawl/commonlid-eval.git@main

src/commonlid/leaderboard/app.py

@@ -410,26 +410,25 @@ def _make_select_handler(
 ) -> Any:
     """Build the row-select callback as a closure over the captured state.
 
-    The callback looks up the clicked row in the *current* table value
-    (passed in via Gradio's event arg) so that switching the scope radio
-    and then clicking a row drills down the row at its post-toggle
-    position, not the row that would have been there before the swap.
+    Uses ``gr.SelectData.row_value`` (Gradio's per-click payload that
+    contains the clicked row as a 1-D list) so the drilldown picks up the
+    *current* table ordering — switching the scope radio and then clicking
+    a row resolves to the row at its post-toggle position. Passing the
+    Dataframe component as an event input would not work: Gradio 6
+    preprocesses Dataframe inputs into ``pandas.DataFrame`` objects, not
+    the ``{"data", "headers"}`` dict we feed in via ``_styled_value``.
 
     Gradio inspects ``__defaults__`` when registering events, and comparing a
     DataFrame default against a type annotation hits an unimplemented arrow
     dtype path. A closure keeps the state out of the function signature.
     """
 
-    def _on_select(table_value: Any, evt: gr.SelectData) -> tuple[str, Any]:
-        if evt.index is None:
+    def _on_select(evt: gr.SelectData) -> tuple[str, Any]:
+        if evt.index is None or not evt.row_value:
             return ("_Click a row to load per-language metrics._", None)
-        row_idx = evt.index[0] if isinstance(evt.index, list | tuple) else evt.index
         try:
-            data = table_value.get("data") if isinstance(table_value, dict) else None
-            if data is None:
-                return ("_Click a row to load per-language metrics._", None)
-            model_id = data[row_idx][0]
-        except (IndexError, KeyError, TypeError):
+            model_id = evt.row_value[0]
+        except (IndexError, TypeError):
             return ("_Could not resolve clicked row._", None)
         per_lang = _per_language_drilldown(snapshot_root, dataset_id, model_id)
         return (
@@ -540,7 +539,6 @@ def build_app(
                     )
                     leaderboard.select(
                         _make_select_handler(dataset_id, snapshot_root),
-                        inputs=[leaderboard],
                         outputs=[drilldown_label, drilldown],
                     )
         gr.Markdown(footer)

tests/unit/test_leaderboard_data.py

@@ -466,6 +466,51 @@ def test_format_table_cov_scope_renders_em_dashes(tmp_path: Path) -> None:
     assert cov_table.iloc[1]["Samples/s"] == "1234.5"
 
 
+def test_row_select_handler_loads_drilldown_from_row_value(tmp_path: Path) -> None:
+    """Clicking a row resolves model_id via ``evt.row_value[0]`` and renders the drilldown.
+
+    Regression: an earlier version pulled the model_id from the Dataframe
+    input, but Gradio 6 preprocesses Dataframe inputs into ``pandas.DataFrame``
+    objects rather than the ``{"data", "headers"}`` dict the app feeds in,
+    so the handler silently returned the "click a row" placeholder.
+    """
+    pytest.importorskip("gradio")
+    from types import SimpleNamespace
+
+    from commonlid.leaderboard.app import _make_select_handler
+
+    per_lang = _per_language_block({"eng": (10, 10, 8, 0.01), "fra": (10, 10, 4, 0.02)})
+    _write_summary(tmp_path, "commonlid", "GlotLID", per_language=per_lang)
+    handler = _make_select_handler("commonlid", tmp_path)
+
+    evt = SimpleNamespace(
+        index=(0, 0),
+        value="GlotLID",
+        row_value=["GlotLID", "80.0", "60.0", "0.10", "2", "1234.5"],
+    )
+    label, payload = handler(evt)
+    assert "GlotLID" in label
+    assert "commonlid" in label
+    assert payload is not None
+    headers = payload["headers"]
+    assert headers[:2] == ["Language", "F1"]
+    languages = [row[0] for row in payload["data"]]
+    assert set(languages) == {"eng", "fra"}
+
+
+def test_row_select_handler_returns_placeholder_when_index_missing(tmp_path: Path) -> None:
+    pytest.importorskip("gradio")
+    from types import SimpleNamespace
+
+    from commonlid.leaderboard.app import _make_select_handler
+
+    handler = _make_select_handler("commonlid", tmp_path)
+    evt = SimpleNamespace(index=None, value=None, row_value=None)
+    label, payload = handler(evt)
+    assert "Click a row" in label
+    assert payload is None
+
+
 def test_scope_radio_change_swaps_table_and_legend(tmp_path: Path) -> None:
     """The scope-change handler returns a fresh styled table + legend Markdown."""
     pytest.importorskip("gradio")