fix sorting

app.py

@@ -1,61 +1,68 @@
-from pathlib import Path
 import ast
 import json
+import os
 import traceback
-import pandas as pd
-import numpy as np
-
-import gradio as gr
-from datasets import Features, Value, load_dataset
 from datetime import datetime
-import os
+from pathlib import Path
 
+import gradio as gr
+import numpy as np
+import pandas as pd
 from about import (
-    PROBLEM_TYPES, TOKEN, CACHE_PATH, API, submissions_repo, results_repo,
-    COLUMN_DISPLAY_NAMES, COUNT_BASED_METRICS, METRIC_GROUPS,
-    METRIC_GROUP_COLORS, COLUMN_TO_GROUP, TRAINING_DATASETS
+    COLUMN_DISPLAY_NAMES,
+    COLUMN_TO_GROUP,
+    COUNT_BASED_METRICS,
+    METRIC_GROUP_COLORS,
+    METRIC_GROUPS,
+    PROBLEM_TYPES,
+    TOKEN,
+    TRAINING_DATASETS,
+    results_repo,
+    submissions_repo,
 )
+from datasets import Features, Value, load_dataset
 
-
-RESULT_FEATURES = Features({
-    "run_name": Value("string"),
-    "timestamp": Value("string"),
-    "n_structures": Value("int64"),
-    "overall_valid_count": Value("int64"),
-    "charge_neutral_count": Value("int64"),
-    "distance_valid_count": Value("int64"),
-    "plausibility_valid_count": Value("int64"),
-    "unique_count": Value("int64"),
-    "novel_count": Value("int64"),
-    "mean_formation_energy": Value("float64"),
-    "formation_energy_std": Value("float64"),
-    "stability_mean_above_hull": Value("float64"),
-    "stability_std_e_above_hull": Value("float64"),
-    "stability_mean_ensemble_std": Value("float64"),
-    "mean_relaxation_RMSD": Value("float64"),
-    "relaxation_RMSE_std": Value("float64"),
-    "stable_count": Value("int64"),
-    "unique_in_stable_count": Value("int64"),
-    "sun_count": Value("int64"),
-    "metastable_count": Value("int64"),
-    "unique_in_metastable_count": Value("int64"),
-    "msun_count": Value("int64"),
-    "JSDistance": Value("float64"),
-    "MMD": Value("float64"),
-    "FrechetDistance": Value("float64"),
-    "element_diversity": Value("float64"),
-    "space_group_diversity": Value("float64"),
-    "site_diversity": Value("float64"),
-    "physical_size_diversity": Value("float64"),
-    "hhi_production_mean": Value("float64"),
-    "hhi_reserve_mean": Value("float64"),
-    "hhi_combined_mean": Value("float64"),
-    "model_name": Value("string"),
-    "relaxed": Value("bool"),
-    "training_set": Value("string"),
-    "paper_link": Value("string"),
-    "notes": Value("string"),
-})
+RESULT_FEATURES = Features(
+    {
+        "run_name": Value("string"),
+        "timestamp": Value("string"),
+        "n_structures": Value("int64"),
+        "overall_valid_count": Value("int64"),
+        "charge_neutral_count": Value("int64"),
+        "distance_valid_count": Value("int64"),
+        "plausibility_valid_count": Value("int64"),
+        "unique_count": Value("int64"),
+        "novel_count": Value("int64"),
+        "mean_formation_energy": Value("float64"),
+        "formation_energy_std": Value("float64"),
+        "stability_mean_above_hull": Value("float64"),
+        "stability_std_e_above_hull": Value("float64"),
+        "stability_mean_ensemble_std": Value("float64"),
+        "mean_relaxation_RMSD": Value("float64"),
+        "relaxation_RMSE_std": Value("float64"),
+        "stable_count": Value("int64"),
+        "unique_in_stable_count": Value("int64"),
+        "sun_count": Value("int64"),
+        "metastable_count": Value("int64"),
+        "unique_in_metastable_count": Value("int64"),
+        "msun_count": Value("int64"),
+        "JSDistance": Value("float64"),
+        "MMD": Value("float64"),
+        "FrechetDistance": Value("float64"),
+        "element_diversity": Value("float64"),
+        "space_group_diversity": Value("float64"),
+        "site_diversity": Value("float64"),
+        "physical_size_diversity": Value("float64"),
+        "hhi_production_mean": Value("float64"),
+        "hhi_reserve_mean": Value("float64"),
+        "hhi_combined_mean": Value("float64"),
+        "model_name": Value("string"),
+        "relaxed": Value("bool"),
+        "training_set": Value("string"),
+        "paper_link": Value("string"),
+        "notes": Value("string"),
+    }
+)
 
 
 def get_leaderboard():
@@ -80,15 +87,46 @@ def get_leaderboard():
     return full_df
 
 
-def format_dataframe(df, show_percentage=False, selected_groups=None, compact_view=True):
-    """Format the dataframe with proper column names and optional percentages."""
+def get_display_datatypes(display_df, original_cols):
+    """Return an explicit per-column datatype list for gr.Dataframe.
+
+    Numeric metric columns are marked as "number" so Gradio's header-click
+    sorting compares them numerically instead of lexicographically. The model
+    column stays "html" (it contains links/symbols) and training_set stays
+    "str".
+    """
+    datatypes = []
+    for i in range(len(display_df.columns)):
+        original_col = original_cols[i] if i < len(original_cols) else None
+        if original_col == "model_name":
+            datatypes.append("html")
+        elif original_col == "training_set":
+            datatypes.append("str")
+        elif original_col == "run_name":
+            datatypes.append("str")
+        else:
+            # Every metric column (counts, percentages, energies, distances,
+            # diversity, HHI, ...) is numeric -> numeric sort.
+            datatypes.append("number")
+    return datatypes
+
+
+def format_dataframe(
+    df, show_percentage=False, selected_groups=None, compact_view=True
+):
+    """Format the dataframe with proper column names and optional percentages.
+
+    Returns a tuple of (styler, datatypes) so the caller can pass an explicit
+    datatype list to gr.Dataframe for correct numeric sorting.
+    """
     if len(df) == 0:
-        return df
+        return df, None
 
     selected_cols = ["model_name"]
 
     if compact_view:
         from about import COMPACT_VIEW_COLUMNS
+
         selected_cols = [col for col in COMPACT_VIEW_COLUMNS if col in df.columns]
     else:
         if "training_set" in df.columns:
@@ -120,7 +158,9 @@ def format_dataframe(df, show_percentage=False, selected_groups=None, compact_vi
             if "paper_link" in df.columns:
                 paper_val = row.get("paper_link", None)
                 if paper_val and isinstance(paper_val, str) and paper_val.strip():
-                    symbols.append(f'<a href="{paper_val.strip()}" target="_blank">📄</a>')
+                    symbols.append(
+                        f'<a href="{paper_val.strip()}" target="_blank">📄</a>'
+                    )
 
             if "relaxed" in df.columns and row.get("relaxed", False):
                 symbols.append("⚡")
@@ -141,6 +181,7 @@ def format_dataframe(df, show_percentage=False, selected_groups=None, compact_vi
         display_df["model_name"] = df.apply(add_model_symbols, axis=1)
 
     if "training_set" in display_df.columns:
+
         def format_training_set(val):
             if val is None or (isinstance(val, float) and np.isnan(val)):
                 return ""
@@ -151,13 +192,20 @@ def format_dataframe(df, show_percentage=False, selected_groups=None, compact_vi
             val = val.replace("'", "").replace('"', "")
             return val
 
-        display_df["training_set"] = display_df["training_set"].apply(format_training_set)
+        display_df["training_set"] = display_df["training_set"].apply(
+            format_training_set
+        )
 
+    # --- Percentage handling -------------------------------------------------
+    # IMPORTANT: keep count-based metrics NUMERIC (do not append a "%" string).
+    # Appending "%" turns the column into strings, which makes Gradio's
+    # header-click sorting lexicographic (e.g. 5.0 -> 3.1 -> 22.6). We instead
+    # store the numeric percentage and move the "%" indicator to the header.
     if show_percentage and "n_structures" in df.columns:
         n_structures = df["n_structures"]
         for col in COUNT_BASED_METRICS:
             if col in display_df.columns:
-                display_df[col] = (df[col] / n_structures * 100).round(1).astype(str) + "%"
+                display_df[col] = (df[col] / n_structures * 100).round(1)
 
     for col in display_df.columns:
         if display_df[col].dtype in ["float64", "float32"]:
@@ -165,14 +213,26 @@ def format_dataframe(df, show_percentage=False, selected_groups=None, compact_vi
 
     baseline_indices = set()
     if "notes" in df.columns:
-        is_baseline = df["notes"].fillna("").str.contains("baseline", case=False, na=False)
+        is_baseline = (
+            df["notes"].fillna("").str.contains("baseline", case=False, na=False)
+        )
         non_baseline_df = display_df[~is_baseline]
         baseline_df = display_df[is_baseline]
         display_df = pd.concat([non_baseline_df, baseline_df]).reset_index(drop=True)
         baseline_indices = set(range(len(non_baseline_df), len(display_df)))
 
-    display_df = display_df.rename(columns=COLUMN_DISPLAY_NAMES)
-    return apply_color_styling(display_df, selected_cols, baseline_indices)
+    datatypes = get_display_datatypes(display_df, selected_cols)
+
+    # Move the "%" indicator into the header so cells stay numeric & sortable.
+    rename_map = dict(COLUMN_DISPLAY_NAMES)
+    if show_percentage:
+        for col in COUNT_BASED_METRICS:
+            if col in rename_map:
+                rename_map[col] = f"{rename_map[col]} (%)"
+
+    display_df = display_df.rename(columns=rename_map)
+    styler = apply_color_styling(display_df, selected_cols, baseline_indices)
+    return styler, datatypes
 
 
 def apply_color_styling(display_df, original_cols, baseline_indices=None):
@@ -214,16 +274,31 @@ def parse_training_set(val):
         return []
 
 
-def update_leaderboard(show_percentage, selected_groups, compact_view, cached_df, sort_by, sort_direction, training_set_filter):
-    """Update the leaderboard based on user selections."""
+def update_leaderboard(
+    show_percentage,
+    selected_groups,
+    compact_view,
+    cached_df,
+    sort_by,
+    sort_direction,
+    training_set_filter,
+):
+    """Update the leaderboard based on user selections.
+
+    Sorting is performed here on the RAW numeric dataframe (before any string
+    formatting), so the "Sort By" dropdown is always numerically correct.
+    """
     df_to_format = cached_df.copy()
 
     ALWAYS_SHOW_MODELS = {"AFLOW", "Alexandria", "OQMD"}
-    if training_set_filter and training_set_filter != "All" and "training_set" in df_to_format.columns:
-        mask = (
-            df_to_format["training_set"].apply(lambda x: training_set_filter in parse_training_set(x))
-            | df_to_format["model_name"].isin(ALWAYS_SHOW_MODELS)
-        )
+    if (
+        training_set_filter
+        and training_set_filter != "All"
+        and "training_set" in df_to_format.columns
+    ):
+        mask = df_to_format["training_set"].apply(
+            lambda x: training_set_filter in parse_training_set(x)
+        ) | df_to_format["model_name"].isin(ALWAYS_SHOW_MODELS)
         df_to_format = df_to_format[mask]
 
     if sort_by and sort_by != "None":
@@ -232,16 +307,33 @@ def update_leaderboard(show_percentage, selected_groups, compact_view, cached_df
 
         if raw_column_name in df_to_format.columns:
             ascending = sort_direction == "Ascending"
-            df_to_format = df_to_format.sort_values(by=raw_column_name, ascending=ascending)
+            df_to_format = df_to_format.sort_values(
+                by=raw_column_name, ascending=ascending
+            )
 
-    return format_dataframe(df_to_format, show_percentage, selected_groups, compact_view)
+    styler, _ = format_dataframe(
+        df_to_format, show_percentage, selected_groups, compact_view
+    )
+    return styler
 
 
 def show_output_box(message):
     return gr.update(value=message, visible=True)
 
 
-def submit_cif_files(model_name, problem_type, cif_files, relaxed, relaxation_settings, training_datasets, training_dataset_other, paper_link, hf_model_link, email, profile: gr.OAuthProfile | None):
+def submit_cif_files(
+    model_name,
+    problem_type,
+    cif_files,
+    relaxed,
+    relaxation_settings,
+    training_datasets,
+    training_dataset_other,
+    paper_link,
+    hf_model_link,
+    email,
+    profile: gr.OAuthProfile | None,
+):
     """Submit structures to the leaderboard."""
     from huggingface_hub import upload_file
 
@@ -269,9 +361,13 @@ def submit_cif_files(model_name, problem_type, cif_files, relaxed, relaxation_se
             "model_name": model_name.strip(),
             "problem_type": problem_type,
             "relaxed": relaxed,
-            "relaxation_settings": relaxation_settings.strip() if relaxed and relaxation_settings else None,
+            "relaxation_settings": relaxation_settings.strip()
+            if relaxed and relaxation_settings
+            else None,
             "training_datasets": training_datasets or [],
-            "training_dataset_other": training_dataset_other.strip() if training_dataset_other else None,
+            "training_dataset_other": training_dataset_other.strip()
+            if training_dataset_other
+            else None,
             "paper_link": paper_link.strip() if paper_link else None,
             "hf_model_link": hf_model_link.strip() if hf_model_link else None,
             "email": email.strip(),
@@ -308,7 +404,11 @@ def submit_cif_files(model_name, problem_type, cif_files, relaxed, relaxation_se
 
         os.unlink(temp_metadata_path)
 
-        return f"Success! Submitted {model_name} for {problem_type} evaluation. Submission ID: {submission_id}", submission_id
+        return (
+            f"Success! Submitted {model_name} for {problem_type} evaluation. "
+            f"Submission ID: {submission_id}",
+            submission_id,
+        )
 
     except Exception as e:
         return f"Error during submission: {str(e)}", None
@@ -317,20 +417,36 @@ def submit_cif_files(model_name, problem_type, cif_files, relaxed, relaxation_se
 def generate_metric_legend_html():
     """Generate HTML table with color-coded metric group legend."""
     metric_details = {
-        "Validity ↑": ("Valid, Charge Neutral, Distance Valid, Plausibility Valid", "↑ Higher is better"),
+        "Validity ↑": (
+            "Valid, Charge Neutral, Distance Valid, Plausibility Valid",
+            "↑ Higher is better",
+        ),
         "Uniqueness & Novelty ↑": ("Unique, Novel", "↑ Higher is better"),
-        "Energy Metrics ↓": ("E Above Hull, Formation Energy, Relaxation RMSD (with std)", "↓ Lower is better"),
+        "Energy Metrics ↓": (
+            "E Above Hull, Formation Energy, Relaxation RMSD (with std)",
+            "↓ Lower is better",
+        ),
         "Stability ↑": ("Stable, Unique in Stable, SUN", "↑ Higher is better"),
-        "Metastability ↑": ("Metastable, Unique in Metastable, MSUN", "↑ Higher is better"),
+        "Metastability ↑": (
+            "Metastable, Unique in Metastable, MSUN",
+            "↑ Higher is better",
+        ),
         "Distribution ↓": ("JS Distance, MMD, FID", "↓ Lower is better"),
-        "Diversity ↑": ("Element, Space Group, Atomic Site, Crystal Size", "↑ Higher is better"),
+        "Diversity ↑": (
+            "Element, Space Group, Atomic Site, Crystal Size",
+            "↑ Higher is better",
+        ),
         "HHI ↓": ("HHI Production, HHI Reserve", "↓ Lower is better"),
     }
 
     html = '<table style="width: 100%; border-collapse: collapse;">'
     html += "<thead><tr>"
-    html += '<th style="border: 1px solid #ddd; padding: 8px; text-align: left;">Color</th>'
-    html += '<th style="border: 1px solid #ddd; padding: 8px; text-align: left;">Group</th>'
+    html += (
+        '<th style="border: 1px solid #ddd; padding: 8px; text-align: left;">Color</th>'
+    )
+    html += (
+        '<th style="border: 1px solid #ddd; padding: 8px; text-align: left;">Group</th>'
+    )
     html += '<th style="border: 1px solid #ddd; padding: 8px; text-align: left;">Metrics</th>'
     html += '<th style="border: 1px solid #ddd; padding: 8px; text-align: left;">Direction</th>'
     html += "</tr></thead><tbody>"
@@ -338,9 +454,12 @@ def generate_metric_legend_html():
     for group, color in METRIC_GROUP_COLORS.items():
         metrics, direction = metric_details.get(group, ("", ""))
         group_name = group.replace("↑", "").replace("↓", "").strip()
-
         html += "<tr>"
-        html += f'<td style="border: 1px solid #ddd; padding: 8px;"><div style="width: 30px; height: 20px; background-color: {color}; border: 1px solid #999;"></div></td>'
+        html += (
+            f'<td style="border: 1px solid #ddd; padding: 8px;">'
+            f'<div style="width: 30px; height: 20px; background-color: {color}; '
+            f'border: 1px solid #999;"></div></td>'
+        )
         html += f'<td style="border: 1px solid #ddd; padding: 8px;"><strong>{group_name}</strong></td>'
         html += f'<td style="border: 1px solid #ddd; padding: 8px;">{metrics}</td>'
         html += f'<td style="border: 1px solid #ddd; padding: 8px;">{direction}</td>'
@@ -352,7 +471,8 @@ def generate_metric_legend_html():
 
 def gradio_interface() -> gr.Blocks:
     with gr.Blocks() as demo:
-        gr.Markdown("""
+        gr.Markdown(
+            """
 # 🔬 LeMat-GenBench: A Unified Benchmark for Generative Models of Crystalline Materials
 
 Generative machine learning models hold great promise for accelerating materials discovery, particularly through the inverse design of inorganic crystals, enabling an unprecedented exploration of chemical space. Yet, the lack of standardized evaluation frameworks makes it difficult to evaluate, compare and further develop these ML models meaningfully.
@@ -360,7 +480,8 @@ Generative machine learning models hold great promise for accelerating materials
 **LeMat-GenBench** introduces a unified benchmark for generative models of crystalline materials, with standardized evaluation metrics for meaningful model comparison, diverse tasks, and this leaderboard to encourage and track community progress.
 
 📄 **Paper**: [arXiv](https://arxiv.org/abs/2512.04562) | 💻 **Code**: [GitHub](https://github.com/LeMaterial/lemat-genbench) | 📧 **Contact**: siddharth.betala [at] entalpic.ai, alexandre.duval [at] entalpic.ai
-""")
+"""
+        )
 
         with gr.Tabs(elem_classes="tab-buttons"):
             with gr.TabItem("🚀 Leaderboard", elem_id="boundary-benchmark-tab-table"):
@@ -421,11 +542,13 @@ Generative machine learning models hold great promise for accelerating materials
 
                     ALWAYS_SHOW_MODELS = {"AFLOW", "Alexandria", "OQMD"}
                     filtered_initial_df = initial_df[
-                        initial_df["training_set"].apply(lambda x: "MP-20" in parse_training_set(x))
+                        initial_df["training_set"].apply(
+                            lambda x: "MP-20" in parse_training_set(x)
+                        )
                         | initial_df["model_name"].isin(ALWAYS_SHOW_MODELS)
                     ]
 
-                    formatted_df = format_dataframe(
+                    formatted_df, formatted_datatypes = format_dataframe(
                         filtered_initial_df,
                         show_percentage=True,
                         selected_groups=list(METRIC_GROUPS.keys()),
@@ -443,8 +566,11 @@ Generative machine learning models hold great promise for accelerating materials
                         value=formatted_df,
                         interactive=False,
                         wrap=True,
-                        datatype=["html"] + [None] * (len(formatted_columns) - 1) if formatted_columns else None,
-                        column_widths=["180px"] + ["160px"] * (len(formatted_columns) - 1) if formatted_columns else None,
+                        datatype=formatted_datatypes if formatted_datatypes else None,
+                        column_widths=["180px"]
+                        + ["160px"] * (len(formatted_columns) - 1)
+                        if formatted_columns
+                        else None,
                         show_fullscreen_button=True,
                     )
 
@@ -458,18 +584,33 @@ Generative machine learning models hold great promise for accelerating materials
                         training_set_filter,
                     ]
 
-                    show_percentage.change(fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table)
-                    selected_groups.change(fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table)
-                    compact_view.change(fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table)
-                    sort_by.change(fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table)
-                    sort_direction.change(fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table)
-                    training_set_filter.change(fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table)
+                    show_percentage.change(
+                        fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table
+                    )
+                    selected_groups.change(
+                        fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table
+                    )
+                    compact_view.change(
+                        fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table
+                    )
+                    sort_by.change(
+                        fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table
+                    )
+                    sort_direction.change(
+                        fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table
+                    )
+                    training_set_filter.change(
+                        fn=update_leaderboard, inputs=inputs, outputs=leaderboard_table
+                    )
 
                 except Exception as e:
                     traceback.print_exc()
-                    gr.Markdown(f"Leaderboard is empty or error loading: {type(e).__name__}: {str(e)}")
+                    gr.Markdown(
+                        f"Leaderboard is empty or error loading: {type(e).__name__}: {str(e)}"
+                    )
 
-                gr.Markdown("""
+                gr.Markdown(
+                    """
 **Symbol Legend:**
 - 📄 Paper available (click to view)
 - ✅ Model output verified
@@ -479,14 +620,17 @@ Generative machine learning models hold great promise for accelerating materials
 
 Verified submissions mean the results came from a model submission rather than a CIF submission.
 
-Models marked as baselines appear below the separator line at the bottom of the table.
-""")
+Models marked as baselines appear below a separator line at the bottom of the table.
+"""
+                )
 
             with gr.TabItem("✉️ Submit", elem_id="boundary-benchmark-tab-table"):
-                gr.Markdown("""
+                gr.Markdown(
+                    """
 # Materials Submission
 Upload a ZIP of CIFs with your structures. To ensure eligibility for the leaderboard, please provide exactly 2,500 representative structures.
-""")
+"""
+                )
                 filename = gr.State(value=None)
 
                 gr.LoginButton()
@@ -516,7 +660,9 @@ Upload a ZIP of CIFs with your structures. To ensure eligibility for the leaderb
                         problem_type = gr.Dropdown(PROBLEM_TYPES, label="Problem Type")
 
                     with gr.Column():
-                        cif_file = gr.File(label="Upload a CSV, a pkl, or a ZIP of CIF files.")
+                        cif_file = gr.File(
+                            label="Upload a CSV, a pkl, or a ZIP of CIF files."
+                        )
                         relaxed = gr.Checkbox(
                             value=False,
                             label="Structures are pre-relaxed",
@@ -548,7 +694,9 @@ Upload a ZIP of CIFs with your structures. To ensure eligibility for the leaderb
                 )
 
                 training_dataset_input.change(
-                    fn=lambda x: gr.update(visible="Others (must specify)" in (x or [])),
+                    fn=lambda x: gr.update(
+                        visible="Others (must specify)" in (x or [])
+                    ),
                     inputs=[training_dataset_input],
                     outputs=[training_dataset_other_input],
                 )
@@ -585,4 +733,4 @@ Upload a ZIP of CIFs with your structures. To ensure eligibility for the leaderb
 
 
 if __name__ == "__main__":
-    gradio_interface().launch(show_error=True)
+    gradio_interface().launch(show_error=True)