enhanced naming of dummy column

app.py

@@ -154,17 +154,15 @@ def load_query(request: gr.Request):  # triggered only once at startup => read q
 
 
 def search_model(df: pd.DataFrame, query: str) -> pd.DataFrame:
-    return df[(df[AutoEvalColumn.model.name].str.contains(query, case=False, na=False))]
-
+    return df[(df[AutoEvalColumn.fullname.name].str.contains(query, case=False, na=False))]
 
 def search_license(df: pd.DataFrame, query: str) -> pd.DataFrame:
     return df[df[AutoEvalColumn.license.name].str.contains(query, case=False, na=False)]
 
-
 def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
     always_here_cols = [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
-    # Use AutoEvalColumn.model.name directly if needed
-    filtered_df = df[always_here_cols + [c for c in COLS if c in df.columns and c in columns]]
+    dummy_col = [AutoEvalColumn.fullname.name]
+    filtered_df = df[always_here_cols + [c for c in COLS if c in df.columns and c in columns] + dummy_col]
     return filtered_df
 
 def filter_queries(query: str, df: pd.DataFrame):
@@ -323,7 +321,9 @@ with demo:
 
             leaderboard_table = gr.components.Dataframe(
                 value=leaderboard_df[
-                    [c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value
+                    [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
+                    + shown_columns.value
+                    + [AutoEvalColumn.fullname.name]
                 ],
                 headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value,
                 datatype=TYPES,

src/display/css_html_js.py

@@ -1,5 +1,11 @@
 custom_css = """
 
+/* Hides the final AutoEvalColumn */
+#llm-benchmark-tab-table table td:last-child,
+#llm-benchmark-tab-table table th:last-child {
+    display: none;
+}
+
 /* Limit the width of the first AutoEvalColumn so that names don't expand too much */
 table td:first-child,
 table th:first-child {

src/display/utils.py

@@ -47,34 +47,37 @@ class ColumnContent:
     dummy: bool = False
 
 
-static_columns = [
-    ["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)],
-    ["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)],
-    ["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)],
-    ["model_type", ColumnContent, ColumnContent("Type", "str", False)],
-    ["architecture", ColumnContent, ColumnContent("Architecture", "str", False)],
-    ["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)],
-    ["precision", ColumnContent, ColumnContent("Precision", "str", False)],
-    ["merged", ColumnContent, ColumnContent("Merged", "bool", False)],
-    ["license", ColumnContent, ColumnContent("Hub License", "str", False)],
-    ["params", ColumnContent, ColumnContent("#Params (B)", "number", False)],
-    ["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)],
-    ["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, hidden=True)],
-    ["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)],
-    ["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)],
-    ["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)],
-]
-
-# Append task specific columns using a comprehension
-task_columns = [[task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)] for task in Tasks]
-
-# Finally, combine them into one list
-auto_eval_column_dict = static_columns + task_columns
+auto_eval_column_dict = []
+# Init
+auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
+auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
+# Scores
+auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
+for task in Tasks:
+    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
+# Model information
+auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
+auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
+auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
+auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
+auto_eval_column_dict.append(["merged", ColumnContent, ColumnContent("Merged", "bool", False)])
+auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
+auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
+auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
+auto_eval_column_dict.append(
+    ["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, hidden=True)]
+)
+auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
+auto_eval_column_dict.append(["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
+auto_eval_column_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
+# Dummy column for the search bar (hidden by the custom CSS)
+auto_eval_column_dict.append(["fullname", ColumnContent, ColumnContent("fullname", "str", False, dummy=True)])
 
 # We use make dataclass to dynamically fill the scores from Tasks
 AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
 
 
+
 @dataclass(frozen=True)
 class EvalQueueColumn:  # Queue column
     model = ColumnContent("model", "markdown", True)
@@ -97,6 +100,7 @@ baseline_row = {
     AutoEvalColumn.truthfulqa.name: 25.0,
     AutoEvalColumn.winogrande.name: 50.0,
     AutoEvalColumn.gsm8k.name: 0.21,
+    AutoEvalColumn.fullname.name: "baseline",
     AutoEvalColumn.model_type.name: "",
     AutoEvalColumn.flagged.name: False,
 }
@@ -121,6 +125,7 @@ human_baseline_row = {
     AutoEvalColumn.truthfulqa.name: 94.0,
     AutoEvalColumn.winogrande.name: 94.0,
     AutoEvalColumn.gsm8k.name: 100,
+    AutoEvalColumn.fullname.name: "human_baseline",
     AutoEvalColumn.model_type.name: "",
     AutoEvalColumn.flagged.name: False,
 }

src/leaderboard/filter_models.py

@@ -128,13 +128,19 @@ DO_NOT_SUBMIT_MODELS = [
     "TigerResearch/tigerbot-70b-chat-v4-4k",  # per authors request
 ]
 
+
 def flag_models(leaderboard_data: list[dict]):
     """Flags models based on external criteria or flagged status."""
     for model_data in leaderboard_data:
-        # Use the primary model name for checking flags
-        flag_key = model_data[AutoEvalColumn.model.name]  # Use the direct model name
+        # Merges and moes are flagged automatically
+        if model_data[AutoEvalColumn.flagged.name]:
+            flag_key = "merged"
+        else:
+            flag_key = model_data[AutoEvalColumn.fullname.name]
 
+        print(f"model check: {flag_key}")
         if flag_key in FLAGGED_MODELS:
+            print(f"Flagged model: {flag_key}")
             issue_num = FLAGGED_MODELS[flag_key].split("/")[-1]
             issue_link = model_hyperlink(
                 FLAGGED_MODELS[flag_key],
@@ -152,8 +158,7 @@ def remove_forbidden_models(leaderboard_data: list[dict]):
     """Removes models from the leaderboard based on the DO_NOT_SUBMIT list."""
     indices_to_remove = []
     for ix, model in enumerate(leaderboard_data):
-        # Use the correct field that now holds the model name
-        if model[AutoEvalColumn.model.name] in DO_NOT_SUBMIT_MODELS:
+        if model[AutoEvalColumn.fullname.name] in DO_NOT_SUBMIT_MODELS:
             indices_to_remove.append(ix)
 
     # Remove the models from the list
@@ -161,6 +166,8 @@ def remove_forbidden_models(leaderboard_data: list[dict]):
         leaderboard_data.pop(ix)
     return leaderboard_data
 
+
 def filter_models_flags(leaderboard_data: list[dict]):
     leaderboard_data = remove_forbidden_models(leaderboard_data)
     flag_models(leaderboard_data)
+

src/leaderboard/read_evals.py

@@ -133,6 +133,7 @@ class EvalResult:
             AutoEvalColumn.weight_type.name: self.weight_type.value.name,
             AutoEvalColumn.architecture.name: self.architecture,
             AutoEvalColumn.model.name: make_clickable_model(self.full_model),
+            AutoEvalColumn.fullname.name: self.full_model,
             AutoEvalColumn.revision.name: self.revision,
             AutoEvalColumn.average.name: average,
             AutoEvalColumn.license.name: self.license,

src/tools/collections.py

@@ -60,7 +60,7 @@ def update_collections(df: DataFrame):
         for size, interval in intervals.items():
             filtered_df = _filter_by_type_and_size(df, model_type, interval)
             best_models = list(
-                filtered_df.sort_values(AutoEvalColumn.average.name, ascending=False)[AutoEvalColumn.model.name][:10]
+                filtered_df.sort_values(AutoEvalColumn.average.name, ascending=False)[AutoEvalColumn.fullname.name][:10]
             )
             print(model_type.value.symbol, size, best_models)
             _add_models_to_collection(collection, best_models, model_type, size)