Initial clone with modifications

.ipynb_checkpoints/app-checkpoint.py

@@ -18,6 +18,364 @@ import plotly.graph_objects as go
 import numpy as np
 
 
+
+
+# === NEW: helper for prompt sensitivity (simple: only NER/REL and 3 prompts) ===
+def calculate_prompt_sensitivity(dataframe, tasks, prompt_ids):
+    """
+    Computes a simple Prompt Sensitivity Index (PSI) over the tasks (NER, REL)
+    using the distribution of 'Best Prompt Id' across the provided prompt_ids.
+    """
+    cv_per_task = []
+    for task in tasks:
+        prompt_col = f"{task} Best Prompt Id"
+        task_accuracies = []
+        for pid in prompt_ids:
+            total = len(dataframe[prompt_col].dropna()) if prompt_col in dataframe.columns else 0
+            count = (dataframe[prompt_col] == pid).sum() if prompt_col in dataframe.columns else 0
+            acc = (count / total * 100) if total > 0 else 0
+            task_accuracies.append(acc)
+        if task_accuracies:
+            mean_acc = np.mean(task_accuracies)
+            std_acc = np.std(task_accuracies)
+            cv_per_task.append((std_acc / mean_acc) if mean_acc > 0 else 0)
+        else:
+            cv_per_task.append(0)
+    mean_cv = np.mean(cv_per_task) if cv_per_task else 0
+    psi = 1.0 if mean_cv >= 0.5 else (mean_cv / 0.5)
+    return psi, mean_cv, cv_per_task
+
+def create_best_model_comparison_table(dataframe, lang: str | None = None, shot: str | None = None):
+    """
+    Table with best overall model per task (NER, REL) and the model with the best prompt score.
+    Applies optional filters:
+      - lang in {EN, IT, SL, SK, GR, PL} or None/"All"
+      - shot in {"0","10"} or None/"All" (mapped to IS_FS False/True)
+    """
+    tasks = ["NER", "REL"]
+    df = dataframe.copy()
+
+    if lang and lang != "All" and "LANG" in df.columns:
+        df = df[df["LANG"] == lang]
+    if shot and shot != "All" and "IS_FS" in df.columns:
+        df = df[df["IS_FS"] == (shot == "10")]
+
+    table_data = {'Task': [], 'Best Overall Model': [], 'CPS': [], 'Best Prompt Model': [], 'Acc.': []}
+
+    for task in tasks:
+        if task not in df.columns or df.empty:
+            continue
+        # Best overall on task
+        max_idx = df[task].idxmax()
+        model_raw = df.loc[max_idx, 'Model']
+        if isinstance(model_raw, str) and '<' in model_raw:
+            match = re.search(r'>([^<]+)<', model_raw)
+            model_name = match.group(1) if match else model_raw
+        else:
+            model_name = str(model_raw)
+        comb_perf_value = df.loc[max_idx, task]
+
+        # Best prompt row for task
+        best_prompt_column = f"{task} Best Prompt"
+        if best_prompt_column in df.columns and df[best_prompt_column].notna().any():
+            best_prompt_idx = df[best_prompt_column].idxmax()
+            best_prompt_model_raw = df.loc[best_prompt_idx, 'Model']
+            if isinstance(best_prompt_model_raw, str) and '<' in best_prompt_model_raw:
+                match = re.search(r'>([^<]+)<', best_prompt_model_raw)
+                best_prompt_model = match.group(1) if match else best_prompt_model_raw
+            else:
+                best_prompt_model = str(best_prompt_model_raw)
+            best_prompt_accuracy = df.loc[best_prompt_idx, best_prompt_column]
+        else:
+            best_prompt_model = "n/a"
+            best_prompt_accuracy = float('nan')
+
+        table_data['Task'].append(task)
+        table_data['Best Overall Model'].append(model_name)
+        table_data['CPS'].append(f"{comb_perf_value:.2f}")
+        table_data['Best Prompt Model'].append(best_prompt_model)
+        table_data['Acc.'].append(f"{best_prompt_accuracy:.2f}" if isinstance(best_prompt_accuracy, (int, float)) else "n/a")
+
+    fig = go.Figure(data=[go.Table(
+        columnwidth=[60, 220, 60, 220, 60],
+        header=dict(
+            values=[f'<b>{col}</b>' for col in table_data.keys()],
+            fill_color=['#2171b5', '#2171b5', '#2171b5', '#4292c6', '#4292c6'],
+            font=dict(color='white', size=12, family='Arial'),
+            align='center', height=30
+        ),
+        cells=dict(
+            values=list(table_data.values()),
+            fill_color=[['#f0f0f0' if i % 2 == 0 else 'white' for i in range(len(table_data['Task']))]],
+            font=dict(color='#2c3e50', size=11, family='Arial'),
+            align=['center', 'left', 'center', 'left', 'center'],
+            height=30
+        )
+    )])
+
+    subtitle = []
+    subtitle.append(lang if (lang and lang != "All") else "All languages")
+    subtitle.append(f"{shot}-shot" if (shot and shot != "All") else "All shots")
+
+    fig.update_layout(
+        title={'text': f"Top Model per Task: CPS & Best Prompt (NER/REL) — {', '.join(subtitle)}",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        font=dict(family="Arial", size=11),
+        height=420, margin=dict(l=20, r=20, t=50, b=80)
+    )
+    return fig
+
+
+
+# === NEW: Best-model comparison table (only NER, REL) ===
+def create_best_model_comparison_table_without_lang(dataframe):
+    """
+    Table with the best overall model per task (NER, REL) and the model that
+    achieves the best score with its own best prompt.
+    """
+    tasks = ["NER", "REL"]
+    table_data = {'Task': [], 'Best Overall Model': [], 'CPS': [], 'Best Prompt Model': [], 'Acc.': []}
+
+    for task in tasks:
+        if task not in dataframe.columns:
+            continue
+
+        # Best overall on the task's combined performance
+        max_idx = dataframe[task].idxmax()
+        model_raw = dataframe.loc[max_idx, 'Model']
+        if isinstance(model_raw, str) and '<' in model_raw:
+            match = re.search(r'>([^<]+)<', model_raw)
+            model_name = match.group(1) if match else model_raw
+        else:
+            model_name = str(model_raw)
+        comb_perf_value = dataframe.loc[max_idx, task]
+
+        # Model with the best prompt for this task
+        best_prompt_column = f"{task} Best Prompt"
+        if best_prompt_column in dataframe.columns:
+            best_prompt_idx = dataframe[best_prompt_column].idxmax()
+            best_prompt_model_raw = dataframe.loc[best_prompt_idx, 'Model']
+            if isinstance(best_prompt_model_raw, str) and '<' in best_prompt_model_raw:
+                match = re.search(r'>([^<]+)<', best_prompt_model_raw)
+                best_prompt_model = match.group(1) if match else best_prompt_model_raw
+            else:
+                best_prompt_model = str(best_prompt_model_raw)
+            best_prompt_accuracy = dataframe.loc[best_prompt_idx, best_prompt_column]
+        else:
+            best_prompt_model = "n/a"
+            best_prompt_accuracy = float('nan')
+
+        table_data['Task'].append(task)
+        table_data['Best Overall Model'].append(model_name)
+        table_data['CPS'].append(f"{comb_perf_value:.2f}")
+        table_data['Best Prompt Model'].append(best_prompt_model)
+        table_data['Acc.'].append(f"{best_prompt_accuracy:.2f}" if isinstance(best_prompt_accuracy, (int, float)) else "n/a")
+
+    fig = go.Figure(data=[go.Table(
+        columnwidth=[60, 220, 60, 220, 60],
+        header=dict(
+            values=[f'<b>{col}</b>' for col in table_data.keys()],
+            fill_color=['#2171b5', '#2171b5', '#2171b5', '#4292c6', '#4292c6'],
+            font=dict(color='white', size=12, family='Arial'),
+            align='center', height=30
+        ),
+        cells=dict(
+            values=list(table_data.values()),
+            fill_color=[['#f0f0f0' if i % 2 == 0 else 'white' for i in range(len(table_data['Task']))]],
+            font=dict(color='#2c3e50', size=11, family='Arial'),
+            align=['center', 'left', 'center', 'left', 'center'],
+            height=30
+        )
+    )])
+    fig.update_layout(
+        title={'text': "Top Model per Task: CPS & Best Prompt (NER/REL)",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        font=dict(family="Arial", size=11),
+        height=420, margin=dict(l=20, r=20, t=50, b=80)
+    )
+    fig.add_annotation(
+        text=("Best Overall Model uses the task's primary metric (CPS). "
+              "Best Prompt Model is the one whose own best prompt yields the highest score."),
+        xref="paper", yref="paper", x=0.5, y=-0.20, showarrow=False,
+        font=dict(size=11, color="gray", family="Arial"), align="center", xanchor="center"
+    )
+    return fig
+
+def create_prompt_heatmap(dataframe, lang: str | None = None, shot: str | None = None):
+    """
+    Heatmap of share (%) of models whose BEST prompt is each pid, for NER/REL with prompts p1..p3.
+    Optional filters:
+      - lang: None or one of EN/IT/SL/SK/GR/PL (None means All)
+      - shot: None or "0"/"10" (None means All) mapped to IS_FS False/True
+    """
+    tasks = ["NER", "REL"]
+
+    df = dataframe.copy()
+    # Language filter
+    if lang and lang != "All" and "LANG" in df.columns:
+        df = df[df["LANG"] == lang]
+    # Shot filter -> IS_FS (10-shot=True, 0-shot=False)
+    if shot and shot != "All" and "IS_FS" in df.columns:
+        df = df[df["IS_FS"] == (shot == "10")]
+
+    # Collect prompt ids present, normalize labels to p1..p3
+    def label_for(pid):
+        if isinstance(pid, str): return pid
+        try: return f"p{int(pid)}"
+        except Exception: return str(pid)
+
+    all_ids = set()
+    for task in tasks:
+        col = f"{task} Best Prompt Id"
+        if col in df.columns:
+            all_ids.update(df[col].dropna().unique())
+    prompt_ids_raw = sorted(list(all_ids), key=lambda x: int(re.sub(r'[^0-9]', '', str(x)) or 0))
+    prompt_ids_raw = [pid for pid in prompt_ids_raw if label_for(pid) in {"p1", "p2", "p3"}] or [1, 2, 3]
+    y_tick_labels = [label_for(pid) for pid in prompt_ids_raw]
+
+    matrix, hovers = [], []
+    for pid in prompt_ids_raw:
+        row, hover_row = [], []
+        for task in tasks:
+            col = f"{task} Best Prompt Id"
+            if col in df.columns and len(df[col].dropna()) > 0:
+                series = df[col].dropna()
+
+                def same_pid(v):
+                    a = re.sub(r'[^0-9]', '', str(v))
+                    b = re.sub(r'[^0-9]', '', str(pid))
+                    return a == b and a != ""
+
+                total = len(series)
+                count = sum(same_pid(v) for v in series)
+                pct = (count / total * 100) if total > 0 else 0
+                row.append(pct)
+                hover_row.append(f"<b>{task} — {label_for(pid)}</b><br>Models: {count}/{total}<br>Percentage: {pct:.1f}%")
+            else:
+                row.append(0); hover_row.append(f"<b>{task} — {label_for(pid)}</b><br>No data")
+        matrix.append(row); hovers.append(hover_row)
+
+    fig = go.Figure(data=go.Heatmap(
+        z=matrix, x=tasks, y=y_tick_labels,
+        colorscale=[[0,'#f7fbff'],[0.2,'#deebf7'],[0.4,'#9ecae1'],[0.6,'#4292c6'],[0.8,'#2171b5'],[1,'#08519c']],
+        text=[[f"{val:.0f}%" if val is not None else "" for val in row] for row in matrix],
+        texttemplate="%{text}", textfont={"size": 11, "family": "Arial"},
+        hovertemplate='%{customdata}<extra></extra>', customdata=hovers,
+        colorbar=dict(title="% Models", ticksuffix="%"),
+        zmin=0, zmax=100
+    ))
+
+    title_parts = []
+    title_parts.append(lang if (lang and lang != "All") else "All languages")
+    title_parts.append(f"{shot}-shot" if (shot and shot != "All") else "All shots")
+    fig.update_layout(
+        title={'text': f"Most Effective Prompts (NER/REL) — {', '.join(title_parts)}",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        xaxis_title="Task", yaxis_title="Prompt",
+        font=dict(family="Arial", size=11), margin=dict(b=100),
+        template="plotly_white", dragmode=False, height=420
+    )
+    fig.update_xaxes(fixedrange=True); fig.update_yaxes(fixedrange=True)
+    return fig
+
+
+# === NEW: Prompt heatmap (only NER, REL; 3 prompts p1, p2, p3) ===
+def create_prompt_heatmap_without_lang(dataframe):
+    """
+    Heatmap of the share of models (in %) whose BEST prompt for the task is each prompt id,
+    for tasks NER and REL, with exactly 3 prompts (p1, p2, p3). It supports columns storing
+    ids as integers (1/2/3) or strings ('p1'/'p2'/'p3').
+    """
+    tasks = ["NER", "REL"]
+
+    # Collect unique prompt ids as they appear (int or 'pX'); restrict to 3 prompts
+    all_ids = set()
+    for task in tasks:
+        col = f"{task} Best Prompt Id"
+        if col in dataframe.columns:
+            all_ids.update(dataframe[col].dropna().unique())
+
+    # Normalize to display labels and preserve the original values as keys
+    def label_for(pid):
+        if isinstance(pid, str):
+            return pid  # e.g., 'p1'
+        try:
+            return f"p{int(pid)}"
+        except Exception:
+            return str(pid)
+
+    prompt_ids_raw = sorted(list(all_ids), key=lambda x: int(re.sub(r'[^0-9]', '', str(x)) or 0))
+    # Optional: hard-limit to p1/p2/p3 if extra noise exists
+    prompt_ids_raw = [pid for pid in prompt_ids_raw if label_for(pid) in {"p1", "p2", "p3"}]
+
+    if not prompt_ids_raw:
+        # Fallback to p1..p3
+        prompt_ids_raw = [1, 2, 3]
+
+    y_tick_labels = [label_for(pid) for pid in prompt_ids_raw]
+
+    matrix, hovers = [], []
+    for pid in prompt_ids_raw:
+        row, hover_row = [], []
+        for task in tasks:
+            col = f"{task} Best Prompt Id"
+            if col in dataframe.columns:
+                series = dataframe[col].dropna()
+                # match values regardless of 'p1' vs 1 vs '1'
+                def same_pid(v):
+                    a = re.sub(r'[^0-9]', '', str(v))
+                    b = re.sub(r'[^0-9]', '', str(pid))
+                    return a == b and a != ""
+                total = len(series)
+                count = sum(same_pid(v) for v in series)
+                pct = (count / total * 100) if total > 0 else 0
+                row.append(pct)
+                hover_row.append(
+                    f"<b>{task} — {label_for(pid)}</b><br>Models: {count}/{total}<br>Percentage: {pct:.1f}%"
+                )
+            else:
+                row.append(0); hover_row.append(f"<b>{task} — {label_for(pid)}</b><br>No data")
+        matrix.append(row)
+        hovers.append(hover_row)
+
+    fig = go.Figure(data=go.Heatmap(
+        z=matrix, x=tasks, y=y_tick_labels,
+        colorscale=[[0,'#f7fbff'],[0.2,'#deebf7'],[0.4,'#9ecae1'],[0.6,'#4292c6'],[0.8,'#2171b5'],[1,'#08519c']],
+        text=[[f"{val:.0f}%" if val is not None else "" for val in row] for row in matrix],
+        texttemplate="%{text}",
+        textfont={"size": 11, "family": "Arial"},
+        hovertemplate='%{customdata}<extra></extra>',
+        customdata=hovers,
+        colorbar=dict(title="% Models", ticksuffix="%"),
+        zmin=0, zmax=100
+    ))
+    fig.update_layout(
+        title={'text': "Most Effective Prompts Across Models (NER/REL)",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        xaxis_title="Task", yaxis_title="Prompt",
+        font=dict(family="Arial", size=11),
+        margin=dict(b=120), template="plotly_white", dragmode=False, height=420
+    )
+
+    # PSI (optional info line)
+    psi, mean_cv, _ = calculate_prompt_sensitivity(
+        dataframe, tasks, prompt_ids_raw
+    )
+    fig.add_annotation(
+        text=f"Prompt Sensitivity (mean CV): {mean_cv:.2f}",
+        xref="paper", yref="paper", x=0.3, y=1.12, showarrow=False,
+        font=dict(size=11, color="#2c3e50", family="Arial")
+    )
+
+    fig.update_xaxes(fixedrange=True); fig.update_yaxes(fixedrange=True)
+    return fig
+
+
+
+
+
+
+
 def mean_of_max_per_field(df):
     """
     Calcola il massimo per ciascun campo e poi la media dei massimi.
@@ -158,33 +516,33 @@ def boxplot_per_task(dataframe=None, baselines=None, references=None):
             ))
 
             # baseline
-            if task in baselines and baselines[task] is not None:
-                fig.add_shape(
-                    type="line",
-                    x0=i - 0.3, x1=i + 0.3,
-                    y0=baselines[task], y1=baselines[task],
-                    line=dict(color="black", width=2, dash="dot"),  # più visibile
-                    xref="x", yref="y"
-                )
-                '''
-                fig.add_annotation(
-                    x=i, y=baselines[task],
-                    text=f"{baselines[task]}%",
-                    showarrow=False,
-                    yshift=10,
-                    font=dict(size=10, color="black")
-                )
-                '''
+            #if task in baselines and baselines[task] is not None:
+                #fig.add_shape(
+                #    type="line",
+                #    x0=i - 0.3, x1=i + 0.3,
+                #    y0=baselines[task], y1=baselines[task],
+                #   line=dict(color="black", width=2, dash="dot"),  # più visibile
+                 #   xref="x", yref="y"
+                #)
+                #'''
+                #fig.add_annotation(
+                    #x=i, y=baselines[task],
+                    #text=f"{baselines[task]}%",
+                    #showarrow=False,
+                    #yshift=10,
+                    #font=dict(size=10, color="black")
+                #)
+                #'''
 
             # reference GPT-4o
-            if task in references and references[task] is not None:
-                fig.add_shape(
-                    type="line",
-                    x0=i - 0.3, x1=i + 0.3,
-                    y0=references[task], y1=references[task],
-                    line=dict(color="red", width=2, dash="dashdot"),
-                    xref="x", yref="y"
-                )
+           # if task in references and references[task] is not None:
+             #   fig.add_shape(
+            #        type="line",
+           #         x0=i - 0.3, x1=i + 0.3,
+           #         y0=references[task], y1=references[task],
+           #         line=dict(color="red", width=2, dash="dashdot"),
+           #         xref="x", yref="y"
+           #     )
 
     fig.update_layout(
         title="Distribution of Model Accuracy by Task",
@@ -268,8 +626,6 @@ def boxplot_prompts_per_task(dataframe, tasks=None):
         y=avg_y,
         name="Avg. Accuracy",
         marker_color="#1f77b4",
-        #hovertemplate="%{y:.2f}%<extra></extra>"
-        #hovertemplate="<b>" + task + "</b><br>Accuracy: %{y:.2f}%<extra></extra>",
     ))
 
     # Barre Best Prompt (rosso)
@@ -278,8 +634,6 @@ def boxplot_prompts_per_task(dataframe, tasks=None):
         y=best_y,
         name="Best Prompt",
         marker_color="#d62728",
-        #hovertemplate="%{y:.2f}%<extra></extra>"
-        #hovertemplate = "<b>" + task + "</b><br>Accuracy: %{y:.2f}%<extra></extra>",
     ))
 
     # Testo sopra barre Best Prompt con ID
@@ -347,7 +701,7 @@ def line_chart(dataframe):
         x=x_true,
         y=y_true,
         mode='markers',
-        name='5-Shot',
+        name='10-Shot',
         marker=dict(
             color='blue',
             size=scale_sizes(x_true)
@@ -418,7 +772,7 @@ def line_chart(dataframe):
     # Caption
     fig.add_annotation(
         text="Accuracy generally rises with #Params, but smaller models <br>"
-             "with 5-shot can outperform larger zero-shot models.",
+             "with 10-shot can outperform larger zero-shot models.",
         xref="paper", yref="paper",
         x=0.5, y=-0.3,  # 👈 centrata
         showarrow=False,
@@ -523,25 +877,13 @@ def init_leaderboard(dataframe, default_selection=None, hidden_columns=None):
     return Leaderboard(
         value=sorted_dataframe,
         datatype=[c.type for c in field_list],
-        #select_columns=SelectColumns(
-        #    default_selection=default_selection or [c.name for c in field_list if c.displayed_by_default],
-        #    cant_deselect=[c.name for c in field_list if c.never_hidden],
-        #    label="Select Columns to Display:",
-        #),
         search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
         hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
         filter_columns=[
-            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
-            #ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Few-Shot Learning (FS)",
-            #             default=[["0️⃣", "0️⃣"]]),
-            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges "),
-
+            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS): "),
+            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges: "),
             ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max = 100, default = [0,100], label="Select the number of parameters (B)"),
         ],
-        #filter_columns=[
-        #    ColumnFilter("IS_FS", type="checkbox", default=False, label="5-Few-Shot")
-        #    #ColumnFilter("FS", type="dropdown", label="5-Few-Shot")
-        #],
         bool_checkboxgroup_label="Evaluation Mode",
         interactive=False,
     )
@@ -620,8 +962,8 @@ def update_task_leaderboard(dataframe, default_selection=None, hidden_columns=No
         search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
         hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
         filter_columns=[
-            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
-            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges "),
+            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS): "),
+            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges:  "),
 
             ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max=100, default=[0, 100],
                          label="Select the number of parameters (B)"),
@@ -687,25 +1029,17 @@ with demo:
             <h1 style="
                 margin: 0 auto; 
                 font-weight: 900; 
-                font-size: 2.5em; 
+                font-size: 5.5em; 
                 letter-spacing: 2px; 
                 text-transform: uppercase; 
+                color: red;
                 background: linear-gradient(90deg, #1f77b4, #00c6ff); 
                 -webkit-background-clip: text; 
                 -webkit-text-fill-color: transparent; 
-                text-shadow: 2px 2px 8px rgba(0,0,0,0.2);
+                text-shadow: 2px 2px 8px rgba(0.2,0,0,0);
             ">
-                EVALITA-LLM Leaderboard
+                ECREAM-LLM Leaderboard
             </h1>
-            <a href="https://huggingface.co/spaces/mii-llm/open_ita_llm_leaderboard" target="_blank" 
-               style="position: absolute; right: 0; display: inline-flex; align-items: center; gap: 6px; text-decoration: none; color: #1f77b4; font-weight: 600;">
-                <!-- Icona stilizzata -->
-                <svg xmlns="http://www.w3.org/2000/svg" width="22" height="22" fill="#1f77b4" viewBox="0 0 24 24">
-                    <path d="M3.9 12a5 5 0 0 1 7.07-7.07l1.41 1.41-1.41 1.41-1.42-1.42a3 3 0 1 0 4.24 4.24l3.54-3.54a5 5 0 0 1-7.07 7.07l-1.41-1.41 1.41-1.41 1.42 1.42z"/>
-                    <path d="M20.1 12a5 5 0 0 1-7.07 7.07l-1.41-1.41 1.41-1.41 1.42 1.42a3 3 0 1 0-4.24-4.24l-3.54 3.54a5 5 0 0 1 7.07-7.07l1.41 1.41-1.41 1.41-1.42-1.42z"/>
-                </svg>
-                Open Italian LLM Leaderboard
-            </a>
         </div>
         """
     )
@@ -715,8 +1049,37 @@ with demo:
     with gr.Row():
         gr.Plot(value=line_chart(LEADERBOARD_DF), elem_id="line-chart")
         gr.Plot(value=boxplot_per_task(LEADERBOARD_DF, BASELINES, REFERENCES), elem_id="boxplot-task")
-        #gr.Plot(value=boxplot_prompts_per_task(LEADERBOARD_DF), elem_id="boxplot-prompt-task")
 
+    # === NEW: second row with the 2 extra plots (NER/REL + p1..p3) ===
+    #with gr.Row():
+        #gr.Plot(value=create_prompt_heatmap(LEADERBOARD_DF), elem_id="prompt-heatmap")
+        #gr.Plot(value=create_best_model_comparison_table(LEADERBOARD_DF), elem_id="best-model-table")
+    # === NEW: gray background wrapper for combos ===
+    with gr.Row(elem_id="filters-wrap"):
+      lang_dd = gr.Dropdown(
+        choices=["All", "EN", "IT", "SL", "SK", "GR", "PL"],
+        value="All", label="Language: ", scale=1
+      )
+      shot_dd = gr.Dropdown(
+        choices=["All", "0", "10"],
+        value="All", label="N-Shot: ", scale=1
+      )
+
+    with gr.Row():
+      heatmap_plot = gr.Plot(value=create_prompt_heatmap(LEADERBOARD_DF, None, None), elem_id="prompt-heatmap")
+      table_plot = gr.Plot(value=create_best_model_comparison_table(LEADERBOARD_DF, None, None), elem_id="best-model-table")
+
+    def _update_both(lang, shot):
+      return (
+        create_prompt_heatmap(LEADERBOARD_DF, None if lang == "All" else lang, None if shot == "All" else shot),
+        create_best_model_comparison_table(LEADERBOARD_DF, None if lang == "All" else lang, None if shot == "All" else shot)
+      )
+
+    lang_dd.change(_update_both, inputs=[lang_dd, shot_dd], outputs=[heatmap_plot, table_plot])
+    shot_dd.change(_update_both, inputs=[lang_dd, shot_dd], outputs=[heatmap_plot, table_plot])
+
+
+  
     with gr.Tabs(elem_classes="tab-buttons") as tabs:
 
         # Main leaderboard tab
@@ -728,41 +1091,11 @@ with demo:
                 hidden_columns=[col for col in LEADERBOARD_DF.columns if col not in ['Rank', 'Size', 'LANG', 'FS', 'Model', "Avg. Comb. Perf. ⬆️", "TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]]
             )
 
-           # gr.HTML(
-             #   f"""
-            #            <div style="
-            #                border: 2px solid #1f77b4;
-            #                border-radius: 10px;
-            #                padding: 10px;
-            #                background-color: #f0f8ff;
-            #                font-weight: bold;
-            #                font-size: 14px;
-            #                display: inline-block;
-            #            ">
-        #                    Theoretical performance of a model that scores the highest on every individual task: <span style="color:#d62728; font-size:18px;">{theoretical_max_combined_perf:.2f}</span>
-          #              </div>
-         # $              """
-          #  )
-
-        '''
-        with gr.TabItem("📈 Charts"):
-            #gr.Plot(value=line_chart(LEADERBOARD_DF), label="Andamento di esempio")
-            #gr.Plot(value=line_chart_interactive_test(), label="Andamento interattivo")
-            gr.Plot(value=line_chart(LEADERBOARD_DF))
-            gr.Plot(value=boxplot_per_task(LEADERBOARD_DF, BASELINES))
-            gr.Plot(value=boxplot_prompts_per_task(LEADERBOARD_DF))
-            gr.Plot(value=barplot_mean_few_minus_zero_shot(LEADERBOARD_DF))
-        '''
 
         # About tab
         with gr.TabItem("📝 About"):
             gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
 
-        # About tab
-        with gr.TabItem("║", interactive=False):
-            gr.Markdown("", elem_classes="markdown-text")
-
-
         # Task-specific leaderboards
         for task, metadata in TASK_METADATA_MULTIPLECHOICE.items():
 
@@ -785,8 +1118,8 @@ with demo:
         for task, metadata in TASK_METADATA_GENERATIVE.items():
             with gr.TabItem(f"{metadata['icon']}{task}"):
                 task_description = TASK_DESCRIPTIONS.get(task, "Description not available.")
-                gr.Markdown(task_description, elem_classes="markdown-text")
-
+                gr.Markdown(task_description, elem_classes="markdown-text1")
+                #print (LEADERBOARD_DF)
                 leaderboard = update_task_leaderboard(
                     LEADERBOARD_DF.rename(columns={f"{task} Prompt Average": "Prompt Average",
                                                    f"{task} Prompt Std": "Prompt Std",
@@ -807,13 +1140,9 @@ with demo:
     with gr.Accordion("📙 Credits", open=False):
         gr.Markdown(
             """
-    **This project has benefited from the following support:**
-
-    - 🧠 **Codebase**: Based on and extended from the Open Italian LLM Leaderboard, developed by **Alessandro Ercolani** and **Samuele Colombo**. We warmly thank them for their invaluable support and guidance in implementing this leaderboard.
-
-    - 💶 **Funding**: Partially supported by the PNRR project **FAIR - Future AI Research (PE00000013)**, under the NRRP MUR program funded by **NextGenerationEU**.
+            ***This project has been funded by the European Union under:
 
-    - 🖥️ **Computation**: We gratefully acknowledge **CINECA** for granting access to the **LEONARDO** supercomputer.  
+                   Horizon Europe eCREAM Project (Grant Agreement No.101057726)  
             """
         )
 

app.py

@@ -18,6 +18,364 @@ import plotly.graph_objects as go
 import numpy as np
 
 
+
+
+# === NEW: helper for prompt sensitivity (simple: only NER/REL and 3 prompts) ===
+def calculate_prompt_sensitivity(dataframe, tasks, prompt_ids):
+    """
+    Computes a simple Prompt Sensitivity Index (PSI) over the tasks (NER, REL)
+    using the distribution of 'Best Prompt Id' across the provided prompt_ids.
+    """
+    cv_per_task = []
+    for task in tasks:
+        prompt_col = f"{task} Best Prompt Id"
+        task_accuracies = []
+        for pid in prompt_ids:
+            total = len(dataframe[prompt_col].dropna()) if prompt_col in dataframe.columns else 0
+            count = (dataframe[prompt_col] == pid).sum() if prompt_col in dataframe.columns else 0
+            acc = (count / total * 100) if total > 0 else 0
+            task_accuracies.append(acc)
+        if task_accuracies:
+            mean_acc = np.mean(task_accuracies)
+            std_acc = np.std(task_accuracies)
+            cv_per_task.append((std_acc / mean_acc) if mean_acc > 0 else 0)
+        else:
+            cv_per_task.append(0)
+    mean_cv = np.mean(cv_per_task) if cv_per_task else 0
+    psi = 1.0 if mean_cv >= 0.5 else (mean_cv / 0.5)
+    return psi, mean_cv, cv_per_task
+
+def create_best_model_comparison_table(dataframe, lang: str | None = None, shot: str | None = None):
+    """
+    Table with best overall model per task (NER, REL) and the model with the best prompt score.
+    Applies optional filters:
+      - lang in {EN, IT, SL, SK, GR, PL} or None/"All"
+      - shot in {"0","10"} or None/"All" (mapped to IS_FS False/True)
+    """
+    tasks = ["NER", "REL"]
+    df = dataframe.copy()
+
+    if lang and lang != "All" and "LANG" in df.columns:
+        df = df[df["LANG"] == lang]
+    if shot and shot != "All" and "IS_FS" in df.columns:
+        df = df[df["IS_FS"] == (shot == "10")]
+
+    table_data = {'Task': [], 'Best Overall Model': [], 'CPS': [], 'Best Prompt Model': [], 'Acc.': []}
+
+    for task in tasks:
+        if task not in df.columns or df.empty:
+            continue
+        # Best overall on task
+        max_idx = df[task].idxmax()
+        model_raw = df.loc[max_idx, 'Model']
+        if isinstance(model_raw, str) and '<' in model_raw:
+            match = re.search(r'>([^<]+)<', model_raw)
+            model_name = match.group(1) if match else model_raw
+        else:
+            model_name = str(model_raw)
+        comb_perf_value = df.loc[max_idx, task]
+
+        # Best prompt row for task
+        best_prompt_column = f"{task} Best Prompt"
+        if best_prompt_column in df.columns and df[best_prompt_column].notna().any():
+            best_prompt_idx = df[best_prompt_column].idxmax()
+            best_prompt_model_raw = df.loc[best_prompt_idx, 'Model']
+            if isinstance(best_prompt_model_raw, str) and '<' in best_prompt_model_raw:
+                match = re.search(r'>([^<]+)<', best_prompt_model_raw)
+                best_prompt_model = match.group(1) if match else best_prompt_model_raw
+            else:
+                best_prompt_model = str(best_prompt_model_raw)
+            best_prompt_accuracy = df.loc[best_prompt_idx, best_prompt_column]
+        else:
+            best_prompt_model = "n/a"
+            best_prompt_accuracy = float('nan')
+
+        table_data['Task'].append(task)
+        table_data['Best Overall Model'].append(model_name)
+        table_data['CPS'].append(f"{comb_perf_value:.2f}")
+        table_data['Best Prompt Model'].append(best_prompt_model)
+        table_data['Acc.'].append(f"{best_prompt_accuracy:.2f}" if isinstance(best_prompt_accuracy, (int, float)) else "n/a")
+
+    fig = go.Figure(data=[go.Table(
+        columnwidth=[60, 220, 60, 220, 60],
+        header=dict(
+            values=[f'<b>{col}</b>' for col in table_data.keys()],
+            fill_color=['#2171b5', '#2171b5', '#2171b5', '#4292c6', '#4292c6'],
+            font=dict(color='white', size=12, family='Arial'),
+            align='center', height=30
+        ),
+        cells=dict(
+            values=list(table_data.values()),
+            fill_color=[['#f0f0f0' if i % 2 == 0 else 'white' for i in range(len(table_data['Task']))]],
+            font=dict(color='#2c3e50', size=11, family='Arial'),
+            align=['center', 'left', 'center', 'left', 'center'],
+            height=30
+        )
+    )])
+
+    subtitle = []
+    subtitle.append(lang if (lang and lang != "All") else "All languages")
+    subtitle.append(f"{shot}-shot" if (shot and shot != "All") else "All shots")
+
+    fig.update_layout(
+        title={'text': f"Top Model per Task: CPS & Best Prompt (NER/REL) — {', '.join(subtitle)}",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        font=dict(family="Arial", size=11),
+        height=420, margin=dict(l=20, r=20, t=50, b=80)
+    )
+    return fig
+
+
+
+# === NEW: Best-model comparison table (only NER, REL) ===
+def create_best_model_comparison_table_without_lang(dataframe):
+    """
+    Table with the best overall model per task (NER, REL) and the model that
+    achieves the best score with its own best prompt.
+    """
+    tasks = ["NER", "REL"]
+    table_data = {'Task': [], 'Best Overall Model': [], 'CPS': [], 'Best Prompt Model': [], 'Acc.': []}
+
+    for task in tasks:
+        if task not in dataframe.columns:
+            continue
+
+        # Best overall on the task's combined performance
+        max_idx = dataframe[task].idxmax()
+        model_raw = dataframe.loc[max_idx, 'Model']
+        if isinstance(model_raw, str) and '<' in model_raw:
+            match = re.search(r'>([^<]+)<', model_raw)
+            model_name = match.group(1) if match else model_raw
+        else:
+            model_name = str(model_raw)
+        comb_perf_value = dataframe.loc[max_idx, task]
+
+        # Model with the best prompt for this task
+        best_prompt_column = f"{task} Best Prompt"
+        if best_prompt_column in dataframe.columns:
+            best_prompt_idx = dataframe[best_prompt_column].idxmax()
+            best_prompt_model_raw = dataframe.loc[best_prompt_idx, 'Model']
+            if isinstance(best_prompt_model_raw, str) and '<' in best_prompt_model_raw:
+                match = re.search(r'>([^<]+)<', best_prompt_model_raw)
+                best_prompt_model = match.group(1) if match else best_prompt_model_raw
+            else:
+                best_prompt_model = str(best_prompt_model_raw)
+            best_prompt_accuracy = dataframe.loc[best_prompt_idx, best_prompt_column]
+        else:
+            best_prompt_model = "n/a"
+            best_prompt_accuracy = float('nan')
+
+        table_data['Task'].append(task)
+        table_data['Best Overall Model'].append(model_name)
+        table_data['CPS'].append(f"{comb_perf_value:.2f}")
+        table_data['Best Prompt Model'].append(best_prompt_model)
+        table_data['Acc.'].append(f"{best_prompt_accuracy:.2f}" if isinstance(best_prompt_accuracy, (int, float)) else "n/a")
+
+    fig = go.Figure(data=[go.Table(
+        columnwidth=[60, 220, 60, 220, 60],
+        header=dict(
+            values=[f'<b>{col}</b>' for col in table_data.keys()],
+            fill_color=['#2171b5', '#2171b5', '#2171b5', '#4292c6', '#4292c6'],
+            font=dict(color='white', size=12, family='Arial'),
+            align='center', height=30
+        ),
+        cells=dict(
+            values=list(table_data.values()),
+            fill_color=[['#f0f0f0' if i % 2 == 0 else 'white' for i in range(len(table_data['Task']))]],
+            font=dict(color='#2c3e50', size=11, family='Arial'),
+            align=['center', 'left', 'center', 'left', 'center'],
+            height=30
+        )
+    )])
+    fig.update_layout(
+        title={'text': "Top Model per Task: CPS & Best Prompt (NER/REL)",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        font=dict(family="Arial", size=11),
+        height=420, margin=dict(l=20, r=20, t=50, b=80)
+    )
+    fig.add_annotation(
+        text=("Best Overall Model uses the task's primary metric (CPS). "
+              "Best Prompt Model is the one whose own best prompt yields the highest score."),
+        xref="paper", yref="paper", x=0.5, y=-0.20, showarrow=False,
+        font=dict(size=11, color="gray", family="Arial"), align="center", xanchor="center"
+    )
+    return fig
+
+def create_prompt_heatmap(dataframe, lang: str | None = None, shot: str | None = None):
+    """
+    Heatmap of share (%) of models whose BEST prompt is each pid, for NER/REL with prompts p1..p3.
+    Optional filters:
+      - lang: None or one of EN/IT/SL/SK/GR/PL (None means All)
+      - shot: None or "0"/"10" (None means All) mapped to IS_FS False/True
+    """
+    tasks = ["NER", "REL"]
+
+    df = dataframe.copy()
+    # Language filter
+    if lang and lang != "All" and "LANG" in df.columns:
+        df = df[df["LANG"] == lang]
+    # Shot filter -> IS_FS (10-shot=True, 0-shot=False)
+    if shot and shot != "All" and "IS_FS" in df.columns:
+        df = df[df["IS_FS"] == (shot == "10")]
+
+    # Collect prompt ids present, normalize labels to p1..p3
+    def label_for(pid):
+        if isinstance(pid, str): return pid
+        try: return f"p{int(pid)}"
+        except Exception: return str(pid)
+
+    all_ids = set()
+    for task in tasks:
+        col = f"{task} Best Prompt Id"
+        if col in df.columns:
+            all_ids.update(df[col].dropna().unique())
+    prompt_ids_raw = sorted(list(all_ids), key=lambda x: int(re.sub(r'[^0-9]', '', str(x)) or 0))
+    prompt_ids_raw = [pid for pid in prompt_ids_raw if label_for(pid) in {"p1", "p2", "p3"}] or [1, 2, 3]
+    y_tick_labels = [label_for(pid) for pid in prompt_ids_raw]
+
+    matrix, hovers = [], []
+    for pid in prompt_ids_raw:
+        row, hover_row = [], []
+        for task in tasks:
+            col = f"{task} Best Prompt Id"
+            if col in df.columns and len(df[col].dropna()) > 0:
+                series = df[col].dropna()
+
+                def same_pid(v):
+                    a = re.sub(r'[^0-9]', '', str(v))
+                    b = re.sub(r'[^0-9]', '', str(pid))
+                    return a == b and a != ""
+
+                total = len(series)
+                count = sum(same_pid(v) for v in series)
+                pct = (count / total * 100) if total > 0 else 0
+                row.append(pct)
+                hover_row.append(f"<b>{task} — {label_for(pid)}</b><br>Models: {count}/{total}<br>Percentage: {pct:.1f}%")
+            else:
+                row.append(0); hover_row.append(f"<b>{task} — {label_for(pid)}</b><br>No data")
+        matrix.append(row); hovers.append(hover_row)
+
+    fig = go.Figure(data=go.Heatmap(
+        z=matrix, x=tasks, y=y_tick_labels,
+        colorscale=[[0,'#f7fbff'],[0.2,'#deebf7'],[0.4,'#9ecae1'],[0.6,'#4292c6'],[0.8,'#2171b5'],[1,'#08519c']],
+        text=[[f"{val:.0f}%" if val is not None else "" for val in row] for row in matrix],
+        texttemplate="%{text}", textfont={"size": 11, "family": "Arial"},
+        hovertemplate='%{customdata}<extra></extra>', customdata=hovers,
+        colorbar=dict(title="% Models", ticksuffix="%"),
+        zmin=0, zmax=100
+    ))
+
+    title_parts = []
+    title_parts.append(lang if (lang and lang != "All") else "All languages")
+    title_parts.append(f"{shot}-shot" if (shot and shot != "All") else "All shots")
+    fig.update_layout(
+        title={'text': f"Most Effective Prompts (NER/REL) — {', '.join(title_parts)}",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        xaxis_title="Task", yaxis_title="Prompt",
+        font=dict(family="Arial", size=11), margin=dict(b=100),
+        template="plotly_white", dragmode=False, height=420
+    )
+    fig.update_xaxes(fixedrange=True); fig.update_yaxes(fixedrange=True)
+    return fig
+
+
+# === NEW: Prompt heatmap (only NER, REL; 3 prompts p1, p2, p3) ===
+def create_prompt_heatmap_without_lang(dataframe):
+    """
+    Heatmap of the share of models (in %) whose BEST prompt for the task is each prompt id,
+    for tasks NER and REL, with exactly 3 prompts (p1, p2, p3). It supports columns storing
+    ids as integers (1/2/3) or strings ('p1'/'p2'/'p3').
+    """
+    tasks = ["NER", "REL"]
+
+    # Collect unique prompt ids as they appear (int or 'pX'); restrict to 3 prompts
+    all_ids = set()
+    for task in tasks:
+        col = f"{task} Best Prompt Id"
+        if col in dataframe.columns:
+            all_ids.update(dataframe[col].dropna().unique())
+
+    # Normalize to display labels and preserve the original values as keys
+    def label_for(pid):
+        if isinstance(pid, str):
+            return pid  # e.g., 'p1'
+        try:
+            return f"p{int(pid)}"
+        except Exception:
+            return str(pid)
+
+    prompt_ids_raw = sorted(list(all_ids), key=lambda x: int(re.sub(r'[^0-9]', '', str(x)) or 0))
+    # Optional: hard-limit to p1/p2/p3 if extra noise exists
+    prompt_ids_raw = [pid for pid in prompt_ids_raw if label_for(pid) in {"p1", "p2", "p3"}]
+
+    if not prompt_ids_raw:
+        # Fallback to p1..p3
+        prompt_ids_raw = [1, 2, 3]
+
+    y_tick_labels = [label_for(pid) for pid in prompt_ids_raw]
+
+    matrix, hovers = [], []
+    for pid in prompt_ids_raw:
+        row, hover_row = [], []
+        for task in tasks:
+            col = f"{task} Best Prompt Id"
+            if col in dataframe.columns:
+                series = dataframe[col].dropna()
+                # match values regardless of 'p1' vs 1 vs '1'
+                def same_pid(v):
+                    a = re.sub(r'[^0-9]', '', str(v))
+                    b = re.sub(r'[^0-9]', '', str(pid))
+                    return a == b and a != ""
+                total = len(series)
+                count = sum(same_pid(v) for v in series)
+                pct = (count / total * 100) if total > 0 else 0
+                row.append(pct)
+                hover_row.append(
+                    f"<b>{task} — {label_for(pid)}</b><br>Models: {count}/{total}<br>Percentage: {pct:.1f}%"
+                )
+            else:
+                row.append(0); hover_row.append(f"<b>{task} — {label_for(pid)}</b><br>No data")
+        matrix.append(row)
+        hovers.append(hover_row)
+
+    fig = go.Figure(data=go.Heatmap(
+        z=matrix, x=tasks, y=y_tick_labels,
+        colorscale=[[0,'#f7fbff'],[0.2,'#deebf7'],[0.4,'#9ecae1'],[0.6,'#4292c6'],[0.8,'#2171b5'],[1,'#08519c']],
+        text=[[f"{val:.0f}%" if val is not None else "" for val in row] for row in matrix],
+        texttemplate="%{text}",
+        textfont={"size": 11, "family": "Arial"},
+        hovertemplate='%{customdata}<extra></extra>',
+        customdata=hovers,
+        colorbar=dict(title="% Models", ticksuffix="%"),
+        zmin=0, zmax=100
+    ))
+    fig.update_layout(
+        title={'text': "Most Effective Prompts Across Models (NER/REL)",
+               'font': {'family': 'Arial', 'size': 14, 'color': '#2c3e50'}},
+        xaxis_title="Task", yaxis_title="Prompt",
+        font=dict(family="Arial", size=11),
+        margin=dict(b=120), template="plotly_white", dragmode=False, height=420
+    )
+
+    # PSI (optional info line)
+    psi, mean_cv, _ = calculate_prompt_sensitivity(
+        dataframe, tasks, prompt_ids_raw
+    )
+    fig.add_annotation(
+        text=f"Prompt Sensitivity (mean CV): {mean_cv:.2f}",
+        xref="paper", yref="paper", x=0.3, y=1.12, showarrow=False,
+        font=dict(size=11, color="#2c3e50", family="Arial")
+    )
+
+    fig.update_xaxes(fixedrange=True); fig.update_yaxes(fixedrange=True)
+    return fig
+
+
+
+
+
+
+
 def mean_of_max_per_field(df):
     """
     Calcola il massimo per ciascun campo e poi la media dei massimi.
@@ -158,33 +516,33 @@ def boxplot_per_task(dataframe=None, baselines=None, references=None):
             ))
 
             # baseline
-            if task in baselines and baselines[task] is not None:
-                fig.add_shape(
-                    type="line",
-                    x0=i - 0.3, x1=i + 0.3,
-                    y0=baselines[task], y1=baselines[task],
-                    line=dict(color="black", width=2, dash="dot"),  # più visibile
-                    xref="x", yref="y"
-                )
-                '''
-                fig.add_annotation(
-                    x=i, y=baselines[task],
-                    text=f"{baselines[task]}%",
-                    showarrow=False,
-                    yshift=10,
-                    font=dict(size=10, color="black")
-                )
-                '''
+            #if task in baselines and baselines[task] is not None:
+                #fig.add_shape(
+                #    type="line",
+                #    x0=i - 0.3, x1=i + 0.3,
+                #    y0=baselines[task], y1=baselines[task],
+                #   line=dict(color="black", width=2, dash="dot"),  # più visibile
+                 #   xref="x", yref="y"
+                #)
+                #'''
+                #fig.add_annotation(
+                    #x=i, y=baselines[task],
+                    #text=f"{baselines[task]}%",
+                    #showarrow=False,
+                    #yshift=10,
+                    #font=dict(size=10, color="black")
+                #)
+                #'''
 
             # reference GPT-4o
-            if task in references and references[task] is not None:
-                fig.add_shape(
-                    type="line",
-                    x0=i - 0.3, x1=i + 0.3,
-                    y0=references[task], y1=references[task],
-                    line=dict(color="red", width=2, dash="dashdot"),
-                    xref="x", yref="y"
-                )
+           # if task in references and references[task] is not None:
+             #   fig.add_shape(
+            #        type="line",
+           #         x0=i - 0.3, x1=i + 0.3,
+           #         y0=references[task], y1=references[task],
+           #         line=dict(color="red", width=2, dash="dashdot"),
+           #         xref="x", yref="y"
+           #     )
 
     fig.update_layout(
         title="Distribution of Model Accuracy by Task",
@@ -268,8 +626,6 @@ def boxplot_prompts_per_task(dataframe, tasks=None):
         y=avg_y,
         name="Avg. Accuracy",
         marker_color="#1f77b4",
-        #hovertemplate="%{y:.2f}%<extra></extra>"
-        #hovertemplate="<b>" + task + "</b><br>Accuracy: %{y:.2f}%<extra></extra>",
     ))
 
     # Barre Best Prompt (rosso)
@@ -278,8 +634,6 @@ def boxplot_prompts_per_task(dataframe, tasks=None):
         y=best_y,
         name="Best Prompt",
         marker_color="#d62728",
-        #hovertemplate="%{y:.2f}%<extra></extra>"
-        #hovertemplate = "<b>" + task + "</b><br>Accuracy: %{y:.2f}%<extra></extra>",
     ))
 
     # Testo sopra barre Best Prompt con ID
@@ -523,25 +877,13 @@ def init_leaderboard(dataframe, default_selection=None, hidden_columns=None):
     return Leaderboard(
         value=sorted_dataframe,
         datatype=[c.type for c in field_list],
-        #select_columns=SelectColumns(
-        #    default_selection=default_selection or [c.name for c in field_list if c.displayed_by_default],
-        #    cant_deselect=[c.name for c in field_list if c.never_hidden],
-        #    label="Select Columns to Display:",
-        #),
         search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
         hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
         filter_columns=[
-            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
-            #ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Few-Shot Learning (FS)",
-            #             default=[["0️⃣", "0️⃣"]]),
-            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges "),
-
+            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS): "),
+            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges: "),
             ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max = 100, default = [0,100], label="Select the number of parameters (B)"),
         ],
-        #filter_columns=[
-        #    ColumnFilter("IS_FS", type="checkbox", default=False, label="5-Few-Shot")
-        #    #ColumnFilter("FS", type="dropdown", label="5-Few-Shot")
-        #],
         bool_checkboxgroup_label="Evaluation Mode",
         interactive=False,
     )
@@ -620,8 +962,8 @@ def update_task_leaderboard(dataframe, default_selection=None, hidden_columns=No
         search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
         hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
         filter_columns=[
-            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
-            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges "),
+            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS): "),
+            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges:  "),
 
             ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max=100, default=[0, 100],
                          label="Select the number of parameters (B)"),
@@ -707,8 +1049,37 @@ with demo:
     with gr.Row():
         gr.Plot(value=line_chart(LEADERBOARD_DF), elem_id="line-chart")
         gr.Plot(value=boxplot_per_task(LEADERBOARD_DF, BASELINES, REFERENCES), elem_id="boxplot-task")
-        #gr.Plot(value=boxplot_prompts_per_task(LEADERBOARD_DF), elem_id="boxplot-prompt-task")
 
+    # === NEW: second row with the 2 extra plots (NER/REL + p1..p3) ===
+    #with gr.Row():
+        #gr.Plot(value=create_prompt_heatmap(LEADERBOARD_DF), elem_id="prompt-heatmap")
+        #gr.Plot(value=create_best_model_comparison_table(LEADERBOARD_DF), elem_id="best-model-table")
+    # === NEW: gray background wrapper for combos ===
+    with gr.Row(elem_id="filters-wrap"):
+      lang_dd = gr.Dropdown(
+        choices=["All", "EN", "IT", "SL", "SK", "GR", "PL"],
+        value="All", label="Language: ", scale=1
+      )
+      shot_dd = gr.Dropdown(
+        choices=["All", "0", "10"],
+        value="All", label="N-Shot: ", scale=1
+      )
+
+    with gr.Row():
+      heatmap_plot = gr.Plot(value=create_prompt_heatmap(LEADERBOARD_DF, None, None), elem_id="prompt-heatmap")
+      table_plot = gr.Plot(value=create_best_model_comparison_table(LEADERBOARD_DF, None, None), elem_id="best-model-table")
+
+    def _update_both(lang, shot):
+      return (
+        create_prompt_heatmap(LEADERBOARD_DF, None if lang == "All" else lang, None if shot == "All" else shot),
+        create_best_model_comparison_table(LEADERBOARD_DF, None if lang == "All" else lang, None if shot == "All" else shot)
+      )
+
+    lang_dd.change(_update_both, inputs=[lang_dd, shot_dd], outputs=[heatmap_plot, table_plot])
+    shot_dd.change(_update_both, inputs=[lang_dd, shot_dd], outputs=[heatmap_plot, table_plot])
+
+
+  
     with gr.Tabs(elem_classes="tab-buttons") as tabs:
 
         # Main leaderboard tab
@@ -720,41 +1091,11 @@ with demo:
                 hidden_columns=[col for col in LEADERBOARD_DF.columns if col not in ['Rank', 'Size', 'LANG', 'FS', 'Model', "Avg. Comb. Perf. ⬆️", "TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]]
             )
 
-           # gr.HTML(
-             #   f"""
-            #            <div style="
-            #                border: 2px solid #1f77b4;
-            #                border-radius: 10px;
-            #                padding: 10px;
-            #                background-color: #f0f8ff;
-            #                font-weight: bold;
-            #                font-size: 14px;
-            #                display: inline-block;
-            #            ">
-        #                    Theoretical performance of a model that scores the highest on every individual task: <span style="color:#d62728; font-size:18px;">{theoretical_max_combined_perf:.2f}</span>
-          #              </div>
-         # $              """
-          #  )
-
-        '''
-        with gr.TabItem("📈 Charts"):
-            #gr.Plot(value=line_chart(LEADERBOARD_DF), label="Andamento di esempio")
-            #gr.Plot(value=line_chart_interactive_test(), label="Andamento interattivo")
-            gr.Plot(value=line_chart(LEADERBOARD_DF))
-            gr.Plot(value=boxplot_per_task(LEADERBOARD_DF, BASELINES))
-            gr.Plot(value=boxplot_prompts_per_task(LEADERBOARD_DF))
-            gr.Plot(value=barplot_mean_few_minus_zero_shot(LEADERBOARD_DF))
-        '''
 
         # About tab
         with gr.TabItem("📝 About"):
             gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
 
-        # About tab
-        #with gr.TabItem("║", interactive=False):
-           # gr.Markdown("", elem_classes="markdown-text")
-
-
         # Task-specific leaderboards
         for task, metadata in TASK_METADATA_MULTIPLECHOICE.items():
 

app_17_10_2025.py

@@ -0,0 +1,815 @@
+import gradio as gr
+from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
+import pandas as pd
+from apscheduler.schedulers.background import BackgroundScheduler
+from huggingface_hub import snapshot_download
+from src.about import CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE
+from src.tasks import TASK_DESCRIPTIONS, MEASURE_DESCRIPTION
+from src.display.css_html_js import custom_css
+from src.display.utils import BENCHMARK_COLS, COLS, EVAL_COLS, EVAL_TYPES, AutoEvalColumn, ModelType, fields, WeightType, Precision
+from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN
+from src.populate import get_evaluation_queue_df, get_leaderboard_df
+from src.submission.submit import add_new_eval
+import random
+import matplotlib.pyplot as plt
+import re
+import plotly.express as px
+import plotly.graph_objects as go
+import numpy as np
+
+
+def mean_of_max_per_field(df):
+    """
+    Calcola il massimo per ciascun campo e poi la media dei massimi.
+
+    Args:
+        df (pd.DataFrame): DataFrame con colonne TE, SA, HS, AT, WIC, FAQ, LS, SU, NER, REL
+
+    Returns:
+        float: media dei valori massimi dei campi
+    """
+    #fields = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
+    fields = ["NER", "REL"]
+    #print(df.columns)
+
+    # Controlla che tutte le colonne esistano nel DataFrame
+    missing = [f for f in fields if f not in df.columns]
+    if missing:
+        raise ValueError(f"Le seguenti colonne mancano nel DataFrame: {missing}")
+
+    # Calcola il massimo per ciascun campo
+    max_values = df[fields].max()
+
+    # Calcola la media dei massimi
+    mean_max = max_values.mean()
+
+    return mean_max
+
+
+def barplot_mean_few_minus_zero_shot(dataframe, tasks=None):
+    if tasks is None:
+        tasks = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
+ 
+    task_means = {}
+
+    for task in tasks:
+        if task not in dataframe.columns:
+            continue
+
+        # Separa few-shot e zero-shot
+        few_shot = dataframe[dataframe['IS_FS'] == True][["Model", task]]
+        zero_shot = dataframe[dataframe['IS_FS'] == False][["Model", task]]
+
+        # Allinea i modelli
+        merged = pd.merge(few_shot, zero_shot, on="Model", suffixes=("_few", "_zero"))
+
+        # Rimuovi righe con valori mancanti
+        merged = merged.dropna(subset=[f"{task}_few", f"{task}_zero"])
+
+        if merged.empty:
+            continue
+
+        # Calcola differenza few - zero
+        diff = merged[f"{task}_few"] - merged[f"{task}_zero"]
+
+        # Calcola la media
+        task_means[task] = diff.mean()
+
+    # Crea barplot
+    fig = go.Figure([go.Bar(
+        x=list(task_means.keys()),
+        y=list(task_means.values()),
+        marker_color="#ff7f0e",
+        text=[f"{v:.2f}" for v in task_means.values()],
+        textposition="outside",
+        hovertemplate="<b>%{x}</b><br>Mean Delta Accuracy: %{y:.2f}%<extra></extra>"
+    )])
+
+    # Linea di riferimento a 0
+    '''
+    fig.add_shape(
+        type="line",
+        x0=-0.5, x1=len(task_means) - 0.5,
+        y0=0, y1=0,
+        line=dict(color="black", width=2, dash="dash"),
+        xref="x", yref="y"
+    )
+    '''
+
+    fig.update_layout(
+        title="Mean Accuracy Difference (Few-shot − Zero-shot) per Task",
+        xaxis_title="",
+        yaxis_title="Mean Delta Combined Performance",
+        template="plotly_white",
+        font=dict(family="Arial", size=13),
+        #margin=dict(b=100)
+    )
+
+    fig.add_annotation(
+        text="10-shot learning generally outperforms zero-shot. <br>"
+             "",
+        xref="paper", yref="paper",
+        x=0, y=-0.2,
+        showarrow=False,
+        font=dict(size=11, color="gray"),
+        align="left"
+    )
+
+    return fig
+
+
+def boxplot_per_task(dataframe=None, baselines=None, references=None):
+
+    #print(dataframe.columns)
+
+    #tasks = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
+    tasks =["NER", "REL"]
+    if dataframe is None:
+        np.random.seed(42)
+        dataframe = pd.DataFrame({
+            task: np.random.uniform(0.4, 0.9, 20) * 100
+            for task in tasks
+        })
+
+    if baselines is None:
+        baselines = {task: np.random.randint(50, 70) for task in tasks}
+
+    colors = ["#1f77b4", "#ff7f0e", "#2ca02c", "#d62728", "#9467bd",
+              "#8c564b", "#e377c2", "#7f7f7f", "#bcbd22", "#17becf"]
+
+    fig = go.Figure()
+
+    for i, task in enumerate(tasks):
+        if task in dataframe.columns:
+            y_data = dataframe[task].dropna().tolist()
+
+            # boxplot
+            fig.add_trace(go.Box(
+                y=y_data,
+                name=task,
+                marker=dict(color=colors[i]),
+                line=dict(color="black", width=2),
+                fillcolor=colors[i],
+                opacity=0.7,
+                hovertemplate="<b>"+task+"</b><br>Accuracy: %{y:.2f}%<extra></extra>",
+                width=0.6,
+                whiskerwidth=0.2,
+                quartilemethod="linear"
+            ))
+
+            # baseline
+            if task in baselines and baselines[task] is not None:
+                fig.add_shape(
+                    type="line",
+                    x0=i - 0.3, x1=i + 0.3,
+                    y0=baselines[task], y1=baselines[task],
+                    line=dict(color="black", width=2, dash="dot"),  # più visibile
+                    xref="x", yref="y"
+                )
+                '''
+                fig.add_annotation(
+                    x=i, y=baselines[task],
+                    text=f"{baselines[task]}%",
+                    showarrow=False,
+                    yshift=10,
+                    font=dict(size=10, color="black")
+                )
+                '''
+
+            # reference GPT-4o
+            if task in references and references[task] is not None:
+                fig.add_shape(
+                    type="line",
+                    x0=i - 0.3, x1=i + 0.3,
+                    y0=references[task], y1=references[task],
+                    line=dict(color="red", width=2, dash="dashdot"),
+                    xref="x", yref="y"
+                )
+
+    fig.update_layout(
+        title="Distribution of Model Accuracy by Task",
+        xaxis_title="Task",
+        yaxis_title="Combined Performance",
+        template="plotly_white",
+        boxmode="group",
+        dragmode=False,
+        font=dict(family="Arial", size=10),
+        margin=dict(b=80),
+    )
+
+    fig.add_annotation(
+        text=(""
+            #"In tasks like TE and SA, models approach the accuracy of supervised <br>"
+            #"models at EVALITA (dashed black line); in NER and REL they remain lower. <br>"
+           # "Dashed red lines show GPT-4o reference results for generative tasks."
+        ),
+        xref="paper", yref="paper",
+        x=0.5, y=-0.30,
+        showarrow=False,
+        font=dict(size=11, color="gray"),
+        align="left"
+    )
+
+    fig.update_yaxes(range=[0, 100], fixedrange=True)
+
+    return fig
+
+# EVALITA results
+BASELINES = {
+    "TE":71.00, "SA": 66.38, "HS": 80.88, "AT": 82.40, "WIC": 85.00,
+    "LS": 38.82, "SU": 38.91, "NER":88.00, "REL": 62.99
+}
+
+# GPT-4o
+REFERENCES = {
+    "NER": 79.11,
+    "REL": 63.32,
+    "LS": 59.25,
+    "SU": 33.04
+
+}
+
+
+def boxplot_prompts_per_task(dataframe, tasks=None):
+    if tasks is None:
+        tasks = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
+
+    # Lista delle colonne da aggiornare
+    cols_to_update = ["REL Best Prompt Id", "NER Best Prompt Id", "SU Best Prompt Id", "LS Best Prompt Id"]
+    # Applichiamo la trasformazione
+    for col in cols_to_update:
+        dataframe[col] = dataframe[col].replace({1: 7, 2: 8})
+
+    fig = go.Figure()
+
+    # Liste per creare una sola voce in legenda per Average e Best
+    avg_x, avg_y = [], []
+    best_x, best_y, best_text = [], [], []
+
+    for task in tasks:
+        avg_col = f"{task} Prompt Average"
+        best_col = f"{task} Best Prompt"
+        best_id_col = f"{task} Best Prompt Id"
+
+        if all(col in dataframe.columns for col in [avg_col, best_col, best_id_col]):
+            avg_value = dataframe[avg_col].mean()
+            avg_x.append(task)
+            avg_y.append(avg_value)
+
+            best_value = dataframe[best_col].mean()
+            best_x.append(task)
+            best_y.append(best_value)
+            best_id = dataframe[best_id_col].mode()[0]  # Most frequent best prompt id
+            best_text.append(f"P:{best_id}")
+
+    # Barre Average Accuracy (azzurro)
+    fig.add_trace(go.Bar(
+        x=avg_x,
+        y=avg_y,
+        name="Avg. Accuracy",
+        marker_color="#1f77b4",
+        #hovertemplate="%{y:.2f}%<extra></extra>"
+        #hovertemplate="<b>" + task + "</b><br>Accuracy: %{y:.2f}%<extra></extra>",
+    ))
+
+    # Barre Best Prompt (rosso)
+    fig.add_trace(go.Bar(
+        x=best_x,
+        y=best_y,
+        name="Best Prompt",
+        marker_color="#d62728",
+        #hovertemplate="%{y:.2f}%<extra></extra>"
+        #hovertemplate = "<b>" + task + "</b><br>Accuracy: %{y:.2f}%<extra></extra>",
+    ))
+
+    # Testo sopra barre Best Prompt con ID
+    for x, y, text in zip(best_x, best_y, best_text):
+        fig.add_annotation(
+            x=x,
+            y=y + 3,  # leggermente sopra la barra
+            text=text,
+            showarrow=False,
+            font=dict(size=12, color="black")
+        )
+
+    fig.update_layout(
+        title= "Prompt Accuracy: Avg vs Best",
+        xaxis_title="Task",
+        yaxis_title="Combined Performance",
+        barmode='group',
+        template="plotly_white",
+        font=dict(family="Arial", size=10),
+        yaxis=dict(range=[0, 100], fixedrange=True)
+    )
+
+    # caption come annotazione separata
+    fig.add_annotation(
+        text="There is no single prompt that performs best across all tasks.<br>"
+             "Different prompts achieve the highest accuracy on different tasks.",
+        xref="paper", yref="paper",
+        x=0.5, y=-0.3,
+        showarrow=False,
+        font=dict(size=11, color="gray"),
+        align="center",
+        xanchor="center"
+    )
+
+    return fig
+
+
+def line_chart(dataframe):
+
+    # Normalizza le dimensioni per avere marker non troppo piccoli né enormi
+    def scale_sizes(values, min_size=8, max_size=30):
+        vmin, vmax = min(values), max(values)
+        return [
+            min_size + (val - vmin) / (vmax - vmin) * (max_size - min_size) if vmax > vmin else (min_size + max_size) / 2
+            for val in values
+        ]
+
+    # dati in base a IS_FS
+    df_true = dataframe[dataframe['IS_FS'] == True]
+    df_false = dataframe[dataframe['IS_FS'] == False]
+
+    # Estrai valori x, y e labels
+    x_true = df_true['#Params (B)'].tolist()
+    y_true = df_true['Avg. Comb. Perf. ⬆️'].tolist()
+    labels_true = [re.search(r'>([^<]+)<', m).group(1) for m in df_true['Model'].tolist()]
+
+    x_false = df_false['#Params (B)'].tolist()
+    y_false = df_false['Avg. Comb. Perf. ⬆️'].tolist()
+    labels_false = [re.search(r'>([^<]+)<', m).group(1) for m in df_false['Model'].tolist()]
+
+    fig = go.Figure()
+
+    # Punti IS_FS=True
+    fig.add_trace(go.Scatter(
+        x=x_true,
+        y=y_true,
+        mode='markers',
+        name='10-Shot',
+        marker=dict(
+            color='blue',
+            size=scale_sizes(x_true)
+        ),
+        hovertemplate='<b>%{customdata}</b><br>#Params: %{x}<br>Performance: %{y}<extra></extra>',
+        customdata=labels_true
+    ))
+
+    # Punti IS_FS=False
+    fig.add_trace(go.Scatter(
+        x=x_false,
+        y=y_false,
+        mode='markers',
+        name='0-Shot',
+        marker=dict(
+            color='red',
+            size=scale_sizes(x_false)
+        ),
+        hovertemplate='<b>%{customdata}</b><br>#Params: %{x}<br>Performance: %{y}<extra></extra>',
+        customdata=labels_false
+    ))
+
+    # Trova il massimo tra tutti i modelli
+    all_y = y_true + y_false
+    all_x = x_true + x_false
+    all_labels = labels_true + labels_false
+    max_idx = all_y.index(max(all_y))
+    max_x = all_x[max_idx]
+    max_y = all_y[max_idx]
+    max_label = all_labels[max_idx]
+
+    # Aggiungi annotazione visibile per il modello migliore
+    fig.add_annotation(
+        x=max_x,
+        y=max_y,
+        #text=f"Top: {max_label} ({max_y:.1f}%)",
+        text=f"{max_label}",
+        showarrow=True,
+        arrowhead=2,
+        arrowsize=1,
+        arrowwidth=2,
+        arrowcolor="black",
+        font=dict(size=11, color="black"),
+        xshift=10,
+        yshift=10,
+        ax = -30, ay = -20,  # sposta la label a sinistra e sopra il punto
+        xanchor = "right"  # allinea la label a destra rispetto al punto
+    )
+
+    fig.update_layout(
+        title="Avg. Combined Performance vs #Params",
+        xaxis_title="#Params (B)",
+        yaxis_title="Avg. Combined Performance",
+        template="plotly_white",
+        hovermode="closest",
+        font=dict(family="Arial", size=10),
+        dragmode=False,
+        xaxis=dict(
+            tickvals=[0, 25, 50, 75, 100, 125],
+            ticktext=["0", "25", "50", "75", "100"]
+        ),
+        yaxis=dict(
+            tickvals=[0, 20, 40, 60, 80, 100],  # 👈 tick fissi
+            range=[0, 100]  # 👈 range bloccato
+        )
+    )
+
+    # Caption
+    fig.add_annotation(
+        text="Accuracy generally rises with #Params, but smaller models <br>"
+             "with 10-shot can outperform larger zero-shot models.",
+        xref="paper", yref="paper",
+        x=0.5, y=-0.3,  # 👈 centrata
+        showarrow=False,
+        font=dict(size=11, color="gray"),
+        align="center",
+        xanchor="center"  # 👈 ancora centrata rispetto al testo
+    )
+
+    fig.update_xaxes(fixedrange=True, rangeslider_visible=False)
+    fig.update_yaxes(fixedrange=True)
+
+    return fig
+
+
+# Define task metadata (icons, names, descriptions)
+TASK_METADATA_MULTIPLECHOICE = {
+    #"TE": {"icon": "📊", "name": "Textual Entailment", "tooltip": ""},
+    #"SA": {"icon": "😃", "name": "Sentiment Analysis", "tooltip": ""},
+    #"HS": {"icon": "⚠️", "name": "Hate Speech", "tooltip": ""},
+    #"AT": {"icon": "🏥", "name": "Admission Test", "tooltip": ""},
+    #"WIC": {"icon": "🔤", "name": "Word in Context", "tooltip": ""},
+    #"FAQ": {"icon": "❓", "name": "Frequently Asked Questions", "tooltip": ""}
+}
+
+# Define task metadata (icons, names, descriptions)
+TASK_METADATA_GENERATIVE = {
+    #"LS": {"icon": "🔄", "name": "Lexical Substitution", "tooltip": ""},
+    #"SU": {"icon": "📝", "name": "Summarization", "tooltip": ""},
+    "NER": {"icon": "🏷️", "name": "Named Entity Recognition", "tooltip": ""},
+    "REL": {"icon": "🔗", "name": "Relation Extraction", "tooltip": ""},
+}
+
+def restart_space():
+    """Restart the Hugging Face space."""
+    API.restart_space(repo_id=REPO_ID)
+
+
+def init_leaderboard(dataframe, default_selection=None, hidden_columns=None):
+    """
+    Initialize and return the leaderboard when it is first loaded or when 'benchmark' is selected.
+    The table is sorted based on the "Avg. Combined Performance" field.
+    """
+    if dataframe is None or dataframe.empty:
+        raise ValueError("Leaderboard DataFrame is empty or None.")
+
+    #print("????????????????????????????????", mean_of_max_per_field(dataframe))
+
+    sorted_dataframe = dataframe.sort_values(by="Avg. Comb. Perf. ⬆️", ascending=False)
+
+    sorted_dataframe = sorted_dataframe.reset_index(drop=True)
+    sorted_dataframe["Rank"] = sorted_dataframe.index + 1
+
+    # Flag per sapere se la medaglia è già stata assegnata per categoria e tipo
+    large_medal_fs_assigned = False
+    medium_medal_fs_assigned = False
+    small_medal_fs_assigned = False
+
+    large_medal_0shot_assigned = False
+    medium_medal_0shot_assigned = False
+    small_medal_0shot_assigned = False
+
+    # Lista temporanea per salvare i nuovi valori della colonna Model
+    new_model_column = []
+
+    for _, row in sorted_dataframe.iterrows():
+        if row['IS_FS']:  # 10-Few-Shot
+            if row["Size"] == "🔵🔵🔵" and not large_medal_fs_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🔵🏆")
+                large_medal_fs_assigned = True
+            elif row["Size"] == "🔵🔵" and not medium_medal_fs_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🏆")
+                medium_medal_fs_assigned = True
+            elif row["Size"] == "🔵" and not small_medal_fs_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🏆")
+                small_medal_fs_assigned = True
+            else:
+                new_model_column.append(row["Model"])
+        else:  # 0-Shot
+            if row["Size"] == "🔵🔵🔵" and not large_medal_0shot_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🔵🎖️")
+                large_medal_0shot_assigned = True
+            elif row["Size"] == "🔵🔵" and not medium_medal_0shot_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🎖️")
+                medium_medal_0shot_assigned = True
+            elif row["Size"] == "🔵" and not small_medal_0shot_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🎖️")
+                small_medal_0shot_assigned = True
+            else:
+                new_model_column.append(row["Model"])
+
+    # Lista delle colonne da aggiornare
+    #cols_to_update = ["REL Best Prompt Id", "NER Best Prompt Id", "SU Best Prompt Id", "LS Best Prompt Id"]
+    # Applichiamo la trasformazione
+    #for col in cols_to_update:
+    #    dataframe[col] = dataframe[col].replace({1: 7, 2: 8})
+
+    # Aggiorna la colonna Model
+    sorted_dataframe["Model"] = new_model_column
+
+    field_list = fields(AutoEvalColumn)
+
+    return Leaderboard(
+        value=sorted_dataframe,
+        datatype=[c.type for c in field_list],
+        #select_columns=SelectColumns(
+        #    default_selection=default_selection or [c.name for c in field_list if c.displayed_by_default],
+        #    cant_deselect=[c.name for c in field_list if c.never_hidden],
+        #    label="Select Columns to Display:",
+        #),
+        search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
+        hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
+        filter_columns=[
+            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
+            #ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Few-Shot Learning (FS)",
+            #             default=[["0️⃣", "0️⃣"]]),
+            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges "),
+
+            ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max = 100, default = [0,100], label="Select the number of parameters (B)"),
+        ],
+        #filter_columns=[
+        #    ColumnFilter("IS_FS", type="checkbox", default=False, label="5-Few-Shot")
+        #    #ColumnFilter("FS", type="dropdown", label="5-Few-Shot")
+        #],
+        bool_checkboxgroup_label="Evaluation Mode",
+        interactive=False,
+    )
+
+def update_task_leaderboard(dataframe, default_selection=None, hidden_columns=None):
+    """
+    Update and return the leaderboard when a specific task is selected.
+    The table is sorted based on the "Combined Performance" field.
+    """
+    if dataframe is None or dataframe.empty:
+        raise ValueError("Leaderboard DataFrame is empty or None.")
+
+    sorted_dataframe = dataframe.sort_values(by="Combined Performance", ascending=False)
+
+    # aggiungo la colonna rank in base alla posizione
+    sorted_dataframe = sorted_dataframe.reset_index(drop=True)
+    sorted_dataframe["Rank"] = sorted_dataframe.index + 1
+
+    # Flag per sapere se la medaglia è già stata assegnata per categoria e tipo
+    large_medal_fs_assigned = False
+    medium_medal_fs_assigned = False
+    small_medal_fs_assigned = False
+
+    large_medal_0shot_assigned = False
+    medium_medal_0shot_assigned = False
+    small_medal_0shot_assigned = False
+
+    # Lista temporanea per salvare i nuovi valori della colonna Model
+    new_model_column = []
+
+    for _, row in sorted_dataframe.iterrows():
+        if row['IS_FS']:  # 5-Few-Shot
+            if row["Size"] == "🔵🔵🔵" and not large_medal_fs_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🔵🏆")
+                large_medal_fs_assigned = True
+            elif row["Size"] == "🔵🔵" and not medium_medal_fs_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🏆")
+                medium_medal_fs_assigned = True
+            elif row["Size"] == "🔵" and not small_medal_fs_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🏆")
+                small_medal_fs_assigned = True
+            else:
+                new_model_column.append(row["Model"])
+        else:  # 0-Shot
+            if row["Size"] == "🔵🔵🔵" and not large_medal_0shot_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🔵🎖️")
+                large_medal_0shot_assigned = True
+            elif row["Size"] == "🔵🔵" and not medium_medal_0shot_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🔵🎖️")
+                medium_medal_0shot_assigned = True
+            elif row["Size"] == "🔵" and not small_medal_0shot_assigned:
+                new_model_column.append(f"{row['Model']} 🔵🎖️")
+                small_medal_0shot_assigned = True
+            else:
+                new_model_column.append(row["Model"])
+
+    # Aggiorna la colonna Model
+    sorted_dataframe["Model"] = new_model_column
+
+    pd.set_option('display.max_colwidth', None)
+    #print("========================", dataframe['Model'])
+
+    #print(sorted_dataframe['Combined Performance'])
+
+    field_list = fields(AutoEvalColumn)
+
+    return Leaderboard(
+        value=sorted_dataframe,
+        #datatype=[c.type for c in field_list],
+        datatype=[c.type for c in field_list] + [int],
+        #select_columns=SelectColumns(
+        #    default_selection=default_selection or [c.name for c in field_list if c.displayed_by_default],
+        #    cant_deselect=[c.name for c in field_list if c.never_hidden],
+        #    label="Select Columns to Display:",
+        #),
+        search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
+        hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
+        filter_columns=[
+            ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
+            ColumnFilter(AutoEvalColumn.LANG.name, type="checkboxgroup", label="Languges "),
+
+            ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max=100, default=[0, 100],
+                         label="Select the number of parameters (B)"),
+        ],
+        bool_checkboxgroup_label="Evaluation Mode",
+        interactive=False
+    )
+
+'''
+# Helper function for leaderboard initialization
+def init_leaderboard(dataframe, default_selection=None, hidden_columns=None):
+    """Initialize and return a leaderboard."""
+    if dataframe is None or dataframe.empty:
+        raise ValueError("Leaderboard DataFrame is empty or None.")
+
+    return Leaderboard(
+        value=dataframe,
+        datatype=[c.type for c in fields(AutoEvalColumn)],
+        select_columns=SelectColumns(
+            default_selection=default_selection or [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
+            cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden],
+            label="Select Columns to Display:",
+        ),
+        search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
+        hide_columns=hidden_columns or [c.name for c in fields(AutoEvalColumn) if c.hidden],
+        filter_columns=[
+            ColumnFilter(AutoEvalColumn.fewshot_type.name, type="checkboxgroup", label="N-Few-Shot Learning (FS)"),
+            ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max=150, label="Select the number of parameters (B)"),
+        ],
+        bool_checkboxgroup_label="Hide models",
+        interactive=False,
+    )
+'''
+
+def download_snapshot(repo, local_dir):
+    """Try to download a snapshot from Hugging Face Hub."""
+    try:
+        print(f"Downloading from {repo} to {local_dir}...")
+        snapshot_download(repo_id=repo, local_dir=local_dir, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN)
+    except Exception as e:
+        print(f"Error downloading {repo}: {e}")
+        restart_space()
+
+
+# Initialize the app by downloading snapshots
+download_snapshot(QUEUE_REPO, EVAL_REQUESTS_PATH)
+download_snapshot(RESULTS_REPO, EVAL_RESULTS_PATH)
+
+# Load leaderboard data
+LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
+finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
+#print(LEADERBOARD_DF.columns.tolist())
+
+theoretical_max_combined_perf = mean_of_max_per_field(LEADERBOARD_DF)
+
+# Prepare the main interface
+demo = gr.Blocks(css=custom_css)
+with demo:
+    #gr.HTML(TITLE)
+    gr.HTML(
+        """
+        <div style="display: flex; align-items: center; position: relative; width: 100%; height: 60px; padding: 10px 0;">
+            <h1 style="
+                margin: 0 auto; 
+                font-weight: 900; 
+                font-size: 5.5em; 
+                letter-spacing: 2px; 
+                text-transform: uppercase; 
+                color: red;
+                background: linear-gradient(90deg, #1f77b4, #00c6ff); 
+                -webkit-background-clip: text; 
+                -webkit-text-fill-color: transparent; 
+                text-shadow: 2px 2px 8px rgba(0.2,0,0,0);
+            ">
+                ECREAM-LLM Leaderboard
+            </h1>
+        </div>
+        """
+    )
+    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
+
+    # ⬇️ QUI aggiungiamo i grafici subito sotto la barra del titolo e sopra le tabs
+    with gr.Row():
+        gr.Plot(value=line_chart(LEADERBOARD_DF), elem_id="line-chart")
+        gr.Plot(value=boxplot_per_task(LEADERBOARD_DF, BASELINES, REFERENCES), elem_id="boxplot-task")
+        #gr.Plot(value=boxplot_prompts_per_task(LEADERBOARD_DF), elem_id="boxplot-prompt-task")
+
+    with gr.Tabs(elem_classes="tab-buttons") as tabs:
+
+        # Main leaderboard tab
+        with gr.TabItem("🏅 Benchmark"):
+
+            leaderboard = init_leaderboard(
+                LEADERBOARD_DF,
+                default_selection=['Rank', 'Size', 'LANG', 'FS', 'Model', "Avg. Comb. Perf. ⬆️", "TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"],
+                hidden_columns=[col for col in LEADERBOARD_DF.columns if col not in ['Rank', 'Size', 'LANG', 'FS', 'Model', "Avg. Comb. Perf. ⬆️", "TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]]
+            )
+
+           # gr.HTML(
+             #   f"""
+            #            <div style="
+            #                border: 2px solid #1f77b4;
+            #                border-radius: 10px;
+            #                padding: 10px;
+            #                background-color: #f0f8ff;
+            #                font-weight: bold;
+            #                font-size: 14px;
+            #                display: inline-block;
+            #            ">
+        #                    Theoretical performance of a model that scores the highest on every individual task: <span style="color:#d62728; font-size:18px;">{theoretical_max_combined_perf:.2f}</span>
+          #              </div>
+         # $              """
+          #  )
+
+        '''
+        with gr.TabItem("📈 Charts"):
+            #gr.Plot(value=line_chart(LEADERBOARD_DF), label="Andamento di esempio")
+            #gr.Plot(value=line_chart_interactive_test(), label="Andamento interattivo")
+            gr.Plot(value=line_chart(LEADERBOARD_DF))
+            gr.Plot(value=boxplot_per_task(LEADERBOARD_DF, BASELINES))
+            gr.Plot(value=boxplot_prompts_per_task(LEADERBOARD_DF))
+            gr.Plot(value=barplot_mean_few_minus_zero_shot(LEADERBOARD_DF))
+        '''
+
+        # About tab
+        with gr.TabItem("📝 About"):
+            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
+
+        # About tab
+        #with gr.TabItem("║", interactive=False):
+           # gr.Markdown("", elem_classes="markdown-text")
+
+
+        # Task-specific leaderboards
+        for task, metadata in TASK_METADATA_MULTIPLECHOICE.items():
+
+            with gr.TabItem(f"{metadata['icon']}{task}"):
+
+                task_description = TASK_DESCRIPTIONS.get(task, "Description not available.")
+                gr.Markdown(task_description, elem_classes="markdown-text")
+
+                leaderboard = update_task_leaderboard(
+                    LEADERBOARD_DF.rename(columns={f"{task} Prompt Average": "Prompt Average", f"{task} Prompt Std": "Prompt Std", f"{task} Best Prompt": "Best Prompt", f"{task} Best Prompt Id": "Best Prompt Id", task: "Combined Performance"}),
+                    default_selection=['Rank', 'Size','LANG', 'FS', 'Model', 'Combined Performance', 'Prompt Average', 'Prompt Std', 'Best Prompt', 'Best Prompt Id'],
+                    hidden_columns=[col for col in LEADERBOARD_DF.columns if col not in ['Rank', 'Size','LANG', 'FS', 'Model', 'Combined Performance', 'Prompt Average', 'Prompt Std', 'Best Prompt', 'Best Prompt Id']]
+                )
+
+        # About tab
+        with gr.TabItem("│", interactive=False):
+            gr.Markdown("", elem_classes="markdown-text")
+
+        # Task-specific leaderboards
+        for task, metadata in TASK_METADATA_GENERATIVE.items():
+            with gr.TabItem(f"{metadata['icon']}{task}"):
+                task_description = TASK_DESCRIPTIONS.get(task, "Description not available.")
+                gr.Markdown(task_description, elem_classes="markdown-text1")
+                #print (LEADERBOARD_DF)
+                leaderboard = update_task_leaderboard(
+                    LEADERBOARD_DF.rename(columns={f"{task} Prompt Average": "Prompt Average",
+                                                   f"{task} Prompt Std": "Prompt Std",
+                                                   f"{task} Best Prompt": "Best Prompt",
+                                                   f"{task} Best Prompt Id": "Best Prompt Id",
+                                                   task: "Combined Performance"}),
+                    default_selection=['Rank', 'Size', 'LANG', 'FS', 'Model', 'Combined Performance', 'Prompt Average', 'Prompt Std', 'Best Prompt',
+                                       'Best Prompt Id'],
+                    hidden_columns=[col for col in LEADERBOARD_DF.columns if
+                                    col not in ['Rank', 'Size','LANG', 'FS', 'Model', 'Combined Performance', 'Prompt Average', 'Prompt Std',
+                                                'Best Prompt', 'Best Prompt Id']]
+                )
+
+    # Citation section
+    with gr.Accordion("📙 Citation", open=False):
+        gr.Textbox(value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, lines=20, elem_id="citation-button", show_copy_button=True)
+
+    with gr.Accordion("📙 Credits", open=False):
+        gr.Markdown(
+            """
+            ***This project has been funded by the European Union under:
+
+                   Horizon Europe eCREAM Project (Grant Agreement No.101057726)  
+            """
+        )
+
+# Background job to restart space
+scheduler = BackgroundScheduler()
+scheduler.add_job(restart_space, "interval", seconds=1800)
+scheduler.start()
+
+# Launch the app with concurrent queueing
+demo.queue(default_concurrency_limit=40).launch(debug=True,  # Enable Gradio debug mode
+        show_error=True)

src/display/css_html_js.py

@@ -1,5 +1,12 @@
 custom_css = """
-
+/* gray background behind the combo row only */
+#filters-wrap {
+  background: #f2f3f5;       /* light gray */
+  border-radius: 12px;       /* optional */
+  padding: 12px 16px;        /* breathing room */
+  margin-bottom: 8px;        /* space before plots */
+  box-sizing: border-box;
+}
 .markdown-text1 {
     font-size: 16px !important;
      max-height: 300px;   /* adjust height as you like */