evalscope 0.16.2__py3-none-any.whl → 0.17.0__py3-none-any.whl

evalscope/app/constants.py

@@ -2,6 +2,7 @@ PLOTLY_THEME = 'plotly_dark'
 REPORT_TOKEN = '@@'
 MODEL_TOKEN = '::'
 DATASET_TOKEN = ', '
+DEFAULT_BAR_WIDTH = 0.2
 LATEX_DELIMITERS = [{
     'left': '$$',
     'right': '$$',

evalscope/app/ui/__init__.py

@@ -0,0 +1,20 @@
+"""
+UI components for the Evalscope dashboard.
+"""
+from .app_ui import create_app_ui
+from .multi_model import MultiModelComponents, create_multi_model_tab
+from .sidebar import SidebarComponents, create_sidebar
+from .single_model import SingleModelComponents, create_single_model_tab
+from .visualization import VisualizationComponents, create_visualization
+
+__all__ = [
+    'create_app_ui',
+    'SidebarComponents',
+    'create_sidebar',
+    'VisualizationComponents',
+    'create_visualization',
+    'SingleModelComponents',
+    'create_single_model_tab',
+    'MultiModelComponents',
+    'create_multi_model_tab',
+]

evalscope/app/ui/app_ui.py

@@ -0,0 +1,52 @@
+"""
+Main UI application for the Evalscope dashboard.
+"""
+import argparse
+import gradio as gr
+
+from evalscope.version import __version__
+from ..utils.localization import get_app_locale
+from .sidebar import create_sidebar
+from .visualization import create_visualization
+
+
+def create_app_ui(args: argparse.Namespace):
+    lang = args.lang
+    locale_dict = get_app_locale(lang)
+
+    with gr.Blocks(title='Evalscope Dashboard') as demo:
+        gr.HTML(f'<h1 style="text-align: left;">{locale_dict["title"]} (v{__version__})</h1>')
+        with gr.Row():
+            with gr.Column(scale=0, min_width=35):
+                toggle_btn = gr.Button('<')
+            with gr.Column(scale=1):
+                gr.HTML(f'<h3 style="text-align: left;">{locale_dict["star_beggar"]}</h3>')
+
+        with gr.Row():
+            with gr.Column(scale=1) as sidebar_column:
+                sidebar_visible = gr.State(True)
+                sidebar = create_sidebar(args.outputs, lang)
+
+            with gr.Column(scale=5):
+                visualization = create_visualization(sidebar, lang)
+
+        @sidebar.load_btn.click(
+            inputs=[sidebar.reports_dropdown],
+            outputs=[visualization.single_model.report_name, visualization.multi_model.multi_report_name])
+        def update_displays(reports_dropdown):
+            if not reports_dropdown:
+                gr.Warning(locale_dict['note'], duration=3)
+                return gr.skip()
+
+            return (
+                gr.update(choices=reports_dropdown, value=reports_dropdown[0]),  # update single model dropdown
+                gr.update(choices=reports_dropdown, value=reports_dropdown)  # update multi model dropdown
+            )
+
+        @toggle_btn.click(inputs=[sidebar_visible], outputs=[sidebar_column, sidebar_visible, toggle_btn])
+        def toggle_sidebar(visible):
+            new_visible = not visible
+            text = '<' if new_visible else '>'
+            return gr.update(visible=new_visible), new_visible, gr.update(value=text)
+
+    return demo

evalscope/app/ui/multi_model.py

@@ -0,0 +1,323 @@
+"""
+Multi model components for the Evalscope dashboard.
+"""
+import gradio as gr
+import os
+import pandas as pd
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from evalscope.report import ReportKey, get_data_frame
+from evalscope.utils.logger import get_logger
+from ..constants import LATEX_DELIMITERS, MODEL_TOKEN, REPORT_TOKEN
+from ..utils.data_utils import (get_acc_report_df, get_compare_report_df, get_model_prediction, get_single_dataset_df,
+                                load_multi_report, load_single_report)
+from ..utils.localization import get_multi_model_locale
+from ..utils.text_utils import convert_markdown_image, process_model_prediction
+from ..utils.visualization import plot_multi_report_radar
+
+if TYPE_CHECKING:
+    from .sidebar import SidebarComponents
+
+logger = get_logger()
+
+
+@dataclass
+class MultiModelComponents:
+    multi_report_name: gr.Dropdown
+
+
+def create_multi_model_tab(sidebar: 'SidebarComponents', lang: str):
+    locale_dict = get_multi_model_locale(lang)
+
+    multi_report_name = gr.Dropdown(label=locale_dict['select_reports'], choices=[], multiselect=True, interactive=True)
+    report_list = gr.State([])
+
+    with gr.Tab(locale_dict['models_overview']):
+        gr.Markdown(locale_dict['model_radar'])
+        radar_plot = gr.Plot(value=None)
+        gr.Markdown(locale_dict['model_scores'])
+        score_table = gr.DataFrame(value=None)
+
+    with gr.Tab(locale_dict['model_comparison_details']):
+        with gr.Row():
+            model_a_select = gr.Dropdown(label=locale_dict['select_model_a'], choices=[], interactive=True)
+            model_b_select = gr.Dropdown(label=locale_dict['select_model_b'], choices=[], interactive=True)
+
+        # States to store selected models' information
+        model_a_report = gr.State(None)
+        model_b_report = gr.State(None)
+        model_a_dir = gr.State(None)
+        model_b_dir = gr.State(None)
+        model_a_name = gr.State(None)
+        model_b_name = gr.State(None)
+
+        dataset_radio = gr.Radio(label=locale_dict['select_dataset'], choices=[], show_label=True, interactive=True)
+
+        gr.Markdown(f"### {locale_dict['model_predictions']}")
+        subset_select = gr.Dropdown(label=locale_dict['select_subset'], choices=[], show_label=True, interactive=True)
+
+        with gr.Row():
+            answer_mode_radio = gr.Radio(
+                label=locale_dict.get('answer_mode'),
+                choices=['All', 'Pass A & B', 'Fail A & B', 'Pass A, Fail B', 'Fail A, Pass B'],
+                value='All',
+                interactive=True)
+            score_threshold = gr.Number(value=0.99, label=locale_dict['score_threshold'], interactive=True)
+
+        data_comparison_df = gr.State(None)
+        filtered_comparison_df = gr.State(None)
+
+        # Statistics row
+        with gr.Row(variant='panel'):
+            with gr.Column():
+                gr.Markdown('### *Counts*')
+                comparison_counts = gr.Markdown('')
+            with gr.Column():
+                page_number = gr.Number(
+                    value=1, label=locale_dict['page'], minimum=1, maximum=1, step=1, interactive=True)
+
+        # Input and Gold answer sections remain at the top
+        with gr.Row(variant='panel'):
+            with gr.Column():
+                gr.Markdown('### *Input*')
+                input_text = gr.Markdown('', elem_id='input_text', latex_delimiters=LATEX_DELIMITERS)
+
+            with gr.Column():
+                gr.Markdown('### *Gold Answer*')
+                gold_text = gr.Markdown('', elem_id='gold_text', latex_delimiters=LATEX_DELIMITERS)
+
+        # Table-like layout for direct comparison
+        with gr.Row():
+            # Headers for the two models
+            with gr.Column(scale=1):
+                gr.Markdown('### *Model A*')
+            with gr.Column(scale=1):
+                gr.Markdown('### *Model B*')
+
+        # Score comparison row
+        with gr.Row():
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Score*')
+                model_a_score = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Score*')
+                model_b_score = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+
+        # Normalized score comparison row
+        with gr.Row():
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Normalized Score*')
+                model_a_nscore = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Normalized Score*')
+                model_b_nscore = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+
+        # Prediction comparison row
+        with gr.Row():
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Prediction*')
+                model_a_pred = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Prediction*')
+                model_b_pred = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+
+        # Generated output comparison row
+        with gr.Row():
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Generated*')
+                model_a_generated = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+            with gr.Column(scale=1, variant='panel'):
+                gr.Markdown('### *Generated*')
+                model_b_generated = gr.Markdown('', latex_delimiters=LATEX_DELIMITERS)
+
+    @multi_report_name.change(
+        inputs=[sidebar.root_path, multi_report_name],
+        outputs=[report_list, radar_plot, score_table, model_a_select, model_b_select])
+    def update_multi_report_data(root_path, multi_report_names):
+        if not multi_report_names:
+            return gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip()
+
+        report_list = load_multi_report(root_path, multi_report_names)
+        report_df, _ = get_acc_report_df(report_list)
+        report_radar_plot = plot_multi_report_radar(report_df)
+        _, styler = get_compare_report_df(report_df)
+
+        # Extract model names for dropdowns
+        model_choices = multi_report_names
+
+        return report_list, report_radar_plot, styler, gr.update(
+            choices=model_choices, value=model_choices[0]), gr.update(
+                choices=model_choices, value=model_choices[1] if len(model_choices) > 1 else None)
+
+    @gr.on(
+        triggers=[model_a_select.change, model_b_select.change],
+        inputs=[sidebar.root_path, model_a_select, model_b_select],
+        outputs=[model_a_report, model_b_report, model_a_dir, model_b_dir, model_a_name, model_b_name, dataset_radio])
+    def update_selected_models(root_path, model_a, model_b):
+        if not model_a or not model_b:
+            return gr.skip()
+
+        # Load individual reports for both models
+        model_a_reports, datasets_a, _ = load_single_report(root_path, model_a)
+        model_b_reports, datasets_b, _ = load_single_report(root_path, model_b)
+
+        # Get common datasets
+        common_datasets = list(set(datasets_a).intersection(set(datasets_b)))
+
+        # Extract work directories and model names
+        model_a_dir = os.path.join(root_path, model_a.split(REPORT_TOKEN)[0])
+        model_b_dir = os.path.join(root_path, model_b.split(REPORT_TOKEN)[0])
+
+        model_a_name = model_a.split(REPORT_TOKEN)[1].split(MODEL_TOKEN)[0]
+        model_b_name = model_b.split(REPORT_TOKEN)[1].split(MODEL_TOKEN)[0]
+
+        return (model_a_reports, model_b_reports, model_a_dir, model_b_dir, model_a_name, model_b_name,
+                gr.update(choices=common_datasets, value=common_datasets[0] if common_datasets else None))
+
+    @gr.on(
+        triggers=[dataset_radio.change],
+        inputs=[dataset_radio, model_a_report, model_b_report],
+        outputs=[subset_select, data_comparison_df])
+    def update_dataset_comparison(dataset_name, model_a_report, model_b_report):
+        if not dataset_name or model_a_report is None or model_b_report is None:
+            return gr.skip()
+
+        # Get dataframes for both models
+        report_df_a = get_data_frame(report_list=model_a_report)
+        data_score_df_a, _ = get_single_dataset_df(report_df_a, dataset_name)
+
+        report_df_b = get_data_frame(report_list=model_b_report)
+        data_score_df_b, _ = get_single_dataset_df(report_df_b, dataset_name)
+
+        # Get subset choices - should be same for both models
+        subsets = data_score_df_a[ReportKey.subset_name].unique().tolist()
+
+        return gr.update(choices=subsets, value=None), None
+
+    @gr.on(
+        triggers=[subset_select.change],
+        inputs=[model_a_dir, model_b_dir, model_a_name, model_b_name, dataset_radio, subset_select],
+        outputs=[data_comparison_df, page_number])
+    def update_comparison_data(model_a_dir, model_b_dir, model_a_name, model_b_name, dataset_name, subset_name):
+        if not subset_name or not dataset_name:
+            return gr.skip()
+
+        # Get predictions for both models
+        df_a = get_model_prediction(model_a_dir, model_a_name, dataset_name, subset_name)
+        df_b = get_model_prediction(model_b_dir, model_b_name, dataset_name, subset_name)
+
+        # Merge dataframes on Input and Gold columns for comparison
+        if df_a is not None and df_b is not None:
+            # Save the Index column if it exists
+            index_a = df_a['Index'].copy()
+            index_b = df_b['Index'].copy()
+
+            df_a = df_a.add_prefix('A_')
+            df_b = df_b.add_prefix('B_')
+
+            # Restore the Index column
+            df_a['Index'] = index_a
+            df_b['Index'] = index_b
+
+            # Merge on Index
+            comparison_df = pd.merge(df_a, df_b, on='Index')
+
+            return comparison_df, 1
+
+        return None, 1
+
+    @gr.on(
+        triggers=[data_comparison_df.change, answer_mode_radio.change, score_threshold.change],
+        inputs=[data_comparison_df, answer_mode_radio, score_threshold],
+        outputs=[filtered_comparison_df, page_number, comparison_counts])
+    def filter_comparison_data(comparison_df, answer_mode, score_threshold):
+        if comparison_df is None:
+            return None, gr.update(value=1, maximum=1), ''
+
+        all_count = len(comparison_df)
+
+        # Apply filtering based on the selected mode and threshold
+        if answer_mode == 'Pass A & B':
+            filtered_df = comparison_df[(comparison_df['A_NScore'] >= score_threshold)
+                                        & (comparison_df['B_NScore'] >= score_threshold)]
+        elif answer_mode == 'Fail A & B':
+            filtered_df = comparison_df[(comparison_df['A_NScore'] < score_threshold)
+                                        & (comparison_df['B_NScore'] < score_threshold)]
+        elif answer_mode == 'Pass A, Fail B':
+            filtered_df = comparison_df[(comparison_df['A_NScore'] >= score_threshold)
+                                        & (comparison_df['B_NScore'] < score_threshold)]
+        elif answer_mode == 'Fail A, Pass B':
+            filtered_df = comparison_df[(comparison_df['A_NScore'] < score_threshold)
+                                        & (comparison_df['B_NScore'] >= score_threshold)]
+        else:  # All
+            filtered_df = comparison_df
+
+        # Count statistics
+        pass_a_count = len(comparison_df[comparison_df['A_NScore'] >= score_threshold])
+        pass_b_count = len(comparison_df[comparison_df['B_NScore'] >= score_threshold])
+        pass_both_count = len(comparison_df[(comparison_df['A_NScore'] >= score_threshold)
+                                            & (comparison_df['B_NScore'] >= score_threshold)])
+        fail_both_count = len(comparison_df[(comparison_df['A_NScore'] < score_threshold)
+                                            & (comparison_df['B_NScore'] < score_threshold)])
+
+        counts_text = (f'### All: {all_count} | Pass A: {pass_a_count} | Pass B: {pass_b_count} | '
+                       f'Pass Both: {pass_both_count} | Fail Both: {fail_both_count}')
+
+        max_page = max(1, len(filtered_df))
+
+        return filtered_df, gr.update(value=1, maximum=max_page), counts_text
+
+    @gr.on(
+        triggers=[filtered_comparison_df.change, page_number.change, model_a_select.change, model_b_select.change],
+        inputs=[
+            filtered_comparison_df, page_number, score_threshold, model_a_select, model_b_select, model_a_name,
+            model_b_name
+        ],
+        outputs=[
+            input_text, gold_text, model_a_generated, model_a_pred, model_a_score, model_a_nscore, model_b_generated,
+            model_b_pred, model_b_score, model_b_nscore
+        ])
+    def update_comparison_display(filtered_df, page_number, score_threshold, model_a_select, model_b_select,
+                                  model_a_name_val, model_b_name_val):
+        if filtered_df is None or len(filtered_df) == 0:
+            return '', '', '', '', '', '', '', '', '', ''
+
+        # Get the row for the current page
+        start = (page_number - 1)
+        if start >= len(filtered_df):
+            return '', '', '', '', '', '', '', '', '', ''
+
+        row = filtered_df.iloc[start]
+
+        # Process common data
+        input_md = process_model_prediction(row['A_Input'])  # Use A's input (same as B's)
+        gold_md = process_model_prediction(row['A_Gold'])  # Use A's gold (same as B's)
+
+        # Process Model A data
+        a_generated_md = process_model_prediction(row['A_Generated'])
+        a_pred_md = convert_markdown_image(process_model_prediction(row['A_Pred']))
+        a_score_md = process_model_prediction(row['A_Score'])
+        a_nscore_val = float(row['A_NScore']) if not pd.isna(row['A_NScore']) else 0.0
+
+        # Process Model B data
+        b_generated_md = process_model_prediction(row['B_Generated'])
+        b_pred_md = convert_markdown_image(process_model_prediction(row['B_Pred']))
+        b_score_md = process_model_prediction(row['B_Score'])
+        b_nscore_val = float(row['B_NScore']) if not pd.isna(row['B_NScore']) else 0.0
+
+        # Apply visual indicators with backgrounds that make differences more obvious
+        if a_nscore_val >= score_threshold:
+            a_nscore_html = f"<div style='background-color:rgb(45,104, 62); padding:10px;'>{a_nscore_val}</div>"
+        else:
+            a_nscore_html = f"<div style='background-color:rgb(151, 31, 44); padding:10px;'>{a_nscore_val}</div>"
+
+        if b_nscore_val >= score_threshold:
+            b_nscore_html = f"<div style='background-color:rgb(45,104, 62); padding:10px;'>{b_nscore_val}</div>"
+        else:
+            b_nscore_html = f"<div style='background-color:rgb(151, 31, 44); padding:10px;'>{b_nscore_val}</div>"
+
+        return (input_md, gold_md, a_generated_md, a_pred_md, a_score_md, a_nscore_html, b_generated_md, b_pred_md,
+                b_score_md, b_nscore_html)
+
+    return MultiModelComponents(multi_report_name=multi_report_name)

evalscope/app/ui/sidebar.py

@@ -0,0 +1,42 @@
+"""
+Sidebar components for the Evalscope dashboard.
+"""
+import gradio as gr
+import os
+from dataclasses import dataclass
+
+from evalscope.utils.logger import get_logger
+from ..utils.data_utils import scan_for_report_folders
+from ..utils.localization import get_sidebar_locale
+
+logger = get_logger()
+
+
+@dataclass
+class SidebarComponents:
+    root_path: gr.Textbox
+    reports_dropdown: gr.Dropdown
+    load_btn: gr.Button
+
+
+def create_sidebar(outputs_dir: str, lang: str):
+    locale_dict = get_sidebar_locale(lang)
+
+    gr.Markdown(f'## {locale_dict["settings"]}')
+    root_path = gr.Textbox(label=locale_dict['report_root_path'], value=outputs_dir, placeholder=outputs_dir, lines=1)
+    reports_dropdown = gr.Dropdown(label=locale_dict['select_reports'], choices=[], multiselect=True, interactive=True)
+    load_btn = gr.Button(locale_dict['load_btn'])
+    gr.Markdown(f'### {locale_dict["note"]}')
+
+    @reports_dropdown.focus(inputs=[root_path], outputs=[reports_dropdown])
+    def update_dropdown_choices(root_path):
+        folders = scan_for_report_folders(root_path)
+        if len(folders) == 0:
+            gr.Warning(locale_dict['warning'], duration=3)
+        return gr.update(choices=folders)
+
+    return SidebarComponents(
+        root_path=root_path,
+        reports_dropdown=reports_dropdown,
+        load_btn=load_btn,
+    )

evalscope/app/ui/single_model.py

@@ -0,0 +1,202 @@
+"""
+Single model components for the Evalscope dashboard.
+"""
+import gradio as gr
+import os
+import pandas as pd
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from evalscope.report import Report, ReportKey, get_data_frame
+from evalscope.utils.logger import get_logger
+from ..constants import DATASET_TOKEN, LATEX_DELIMITERS, MODEL_TOKEN, REPORT_TOKEN
+from ..utils.data_utils import (get_acc_report_df, get_model_prediction, get_report_analysis, get_single_dataset_df,
+                                load_single_report)
+from ..utils.localization import get_single_model_locale
+from ..utils.text_utils import convert_markdown_image, process_json_content, process_model_prediction
+from ..utils.visualization import plot_single_dataset_scores, plot_single_report_scores, plot_single_report_sunburst
+
+if TYPE_CHECKING:
+    from .sidebar import SidebarComponents
+
+logger = get_logger()
+
+
+@dataclass
+class SingleModelComponents:
+    report_name: gr.Dropdown
+
+
+def create_single_model_tab(sidebar: 'SidebarComponents', lang: str):
+    locale_dict = get_single_model_locale(lang)
+
+    # Update the UI components with localized labels
+    report_name = gr.Dropdown(label=locale_dict['select_report'], choices=[], interactive=True)
+    work_dir = gr.State(None)
+    model_name = gr.State(None)
+
+    with gr.Accordion(locale_dict['task_config'], open=False):
+        task_config = gr.JSON(value=None)
+
+    report_list = gr.State([])
+
+    with gr.Tab(locale_dict['datasets_overview']):
+        gr.Markdown(f'### {locale_dict["dataset_components"]}')
+        sunburst_plot = gr.Plot(value=None, scale=1, label=locale_dict['dataset_components'])
+        gr.Markdown(f'### {locale_dict["dataset_scores"]}')
+        score_plot = gr.Plot(value=None, scale=1, label=locale_dict['dataset_scores'])
+        gr.Markdown(f'### {locale_dict["dataset_scores_table"]}')
+        score_table = gr.DataFrame(value=None)
+
+    with gr.Tab(locale_dict['dataset_details']):
+        dataset_radio = gr.Radio(label=locale_dict['select_dataset'], choices=[], show_label=True, interactive=True)
+        # show dataset details
+        with gr.Accordion(locale_dict['report_analysis'], open=True):
+            report_analysis = gr.Markdown(value='N/A')
+        gr.Markdown(f'### {locale_dict["dataset_scores"]}')
+        dataset_plot = gr.Plot(value=None, scale=1, label=locale_dict['dataset_scores'])
+        gr.Markdown(f'### {locale_dict["dataset_scores_table"]}')
+        dataset_table = gr.DataFrame(value=None)
+
+        gr.Markdown(f'### {locale_dict["model_prediction"]}')
+        subset_select = gr.Dropdown(label=locale_dict['select_subset'], choices=[], show_label=True, interactive=True)
+
+        with gr.Row():
+            answer_mode_radio = gr.Radio(
+                label=locale_dict['answer_mode'], choices=['All', 'Pass', 'Fail'], value='All', interactive=True)
+            score_threshold = gr.Number(value=0.99, label=locale_dict['score_threshold'], interactive=True)
+
+        data_review_df = gr.State(None)
+        filtered_review_df = gr.State(None)
+
+        # show statistics
+        with gr.Row(variant='panel'):
+            with gr.Column():
+                gr.Markdown('### *Counts*')
+                answer_mode_counts = gr.Markdown('')
+            with gr.Column():
+                page_number = gr.Number(
+                    value=1, label=locale_dict['page'], minimum=1, maximum=1, step=1, interactive=True)
+
+        # show data review table
+        with gr.Row(variant='panel'):
+            with gr.Column():
+                gr.Markdown('### *Score*')
+                score_text = gr.Code('', elem_id='score_text', language='json', wrap_lines=False)
+            with gr.Column():
+                gr.Markdown('### *Normalized Score*')
+                nscore = gr.Markdown('', elem_id='score_text', latex_delimiters=LATEX_DELIMITERS)
+
+        with gr.Row(variant='panel'):
+            with gr.Column():
+                gr.Markdown('### *Gold*')
+                gold_text = gr.Markdown('', elem_id='gold_text', latex_delimiters=LATEX_DELIMITERS)
+            with gr.Column():
+                gr.Markdown('### *Pred*')
+                pred_text = gr.Markdown('', elem_id='pred_text', latex_delimiters=LATEX_DELIMITERS)
+
+        with gr.Row(variant='panel'):
+            with gr.Column():
+                gr.Markdown('### *Input*')
+                input_text = gr.Code('', elem_id='input_text', language='json', wrap_lines=False)
+            with gr.Column():
+                gr.Markdown('### *Generated*')
+                generated_text = gr.Markdown('', elem_id='generated_text', latex_delimiters=LATEX_DELIMITERS)
+
+    @report_name.change(
+        inputs=[sidebar.root_path, report_name],
+        outputs=[report_list, task_config, dataset_radio, work_dir, model_name])
+    def update_single_report_data(root_path, report_name):
+        report_list, datasets, task_cfg = load_single_report(root_path, report_name)
+        work_dir = os.path.join(root_path, report_name.split(REPORT_TOKEN)[0])
+        model_name = report_name.split(REPORT_TOKEN)[1].split(MODEL_TOKEN)[0]
+        return (report_list, task_cfg, gr.update(choices=datasets, value=datasets[0]), work_dir, model_name)
+
+    @report_list.change(inputs=[report_list], outputs=[score_plot, score_table, sunburst_plot])
+    def update_single_report_score(report_list):
+        report_score_df, styler = get_acc_report_df(report_list)
+        report_score_plot = plot_single_report_scores(report_score_df)
+        report_sunburst_plot = plot_single_report_sunburst(report_list)
+        return report_score_plot, styler, report_sunburst_plot
+
+    @gr.on(
+        triggers=[dataset_radio.change, report_list.change],
+        inputs=[dataset_radio, report_list],
+        outputs=[dataset_plot, dataset_table, subset_select, data_review_df, report_analysis])
+    def update_single_report_dataset(dataset_name, report_list):
+        logger.debug(f'Updating single report dataset: {dataset_name}')
+        report_df = get_data_frame(report_list=report_list)
+        analysis = get_report_analysis(report_list, dataset_name)
+        data_score_df, styler = get_single_dataset_df(report_df, dataset_name)
+        data_score_plot = plot_single_dataset_scores(data_score_df)
+        subsets = data_score_df[ReportKey.subset_name].unique().tolist()
+        logger.debug(f'subsets: {subsets}')
+        return data_score_plot, styler, gr.update(choices=subsets, value=None), None, analysis
+
+    @gr.on(
+        triggers=[subset_select.change],
+        inputs=[work_dir, model_name, dataset_radio, subset_select],
+        outputs=[data_review_df, page_number])
+    def update_single_report_subset(work_dir, model_name, dataset_name, subset_name):
+        if not subset_name:
+            return gr.skip()
+        data_review_df = get_model_prediction(work_dir, model_name, dataset_name, subset_name)
+        return data_review_df, 1
+
+    @gr.on(
+        triggers=[data_review_df.change, answer_mode_radio.change, score_threshold.change],
+        inputs=[data_review_df, answer_mode_radio, score_threshold],
+        outputs=[filtered_review_df, page_number, answer_mode_counts])
+    def filter_data(data_review_df, answer_mode, score_threshold):
+        if data_review_df is None:
+            return None, gr.update(value=1, maximum=1), ''
+
+        all_count = len(data_review_df)
+        pass_df = data_review_df[data_review_df['NScore'] >= score_threshold]
+        pass_count = len(pass_df)
+        fail_count = all_count - pass_count
+
+        counts_text = f'### All: {all_count} | Pass: {pass_count} | Fail: {fail_count}'
+
+        if answer_mode == 'Pass':
+            filtered_df = pass_df
+        elif answer_mode == 'Fail':
+            filtered_df = data_review_df[data_review_df['NScore'] < score_threshold]
+        else:
+            filtered_df = data_review_df
+
+        max_page = max(1, len(filtered_df))
+
+        return (filtered_df, gr.update(value=1, maximum=max_page), counts_text)
+
+    @gr.on(
+        triggers=[filtered_review_df.change, page_number.change],
+        inputs=[filtered_review_df, page_number, score_threshold],
+        outputs=[input_text, generated_text, gold_text, pred_text, score_text, nscore])
+    def update_table_components(filtered_df, page_number, score_threshold):
+        if filtered_df is None or len(filtered_df) == 0:
+            return '', '', '', '', '', ''
+
+        # Get single row data for the current page
+        start = (page_number - 1)
+        if start >= len(filtered_df):
+            return '', '', '', '', '', ''
+
+        row = filtered_df.iloc[start]
+
+        # Process the data for display
+        input_md = process_json_content(row['Input'])
+        generated_md = process_model_prediction(row['Generated'])
+        gold_md = process_model_prediction(row['Gold'])
+        pred_md = convert_markdown_image(process_model_prediction(row['Pred']))
+        score_md = process_json_content(row['Score'])
+        nscore_val = float(row['NScore']) if not pd.isna(row['NScore']) else 0.0
+
+        if nscore_val >= score_threshold:
+            nscore_val = f'<div style="background-color:rgb(45,104, 62); padding:10px;">{nscore_val}</div>'
+        else:
+            nscore_val = f'<div style="background-color:rgb(151, 31, 44); padding:10px;">{nscore_val}</div>'
+
+        return input_md, generated_md, gold_md, pred_md, score_md, nscore_val
+
+    return SingleModelComponents(report_name=report_name)