import gradio as gr import pandas as pd import plotly.graph_objects as go import os import base64 import re from huggingface_hub import HfApi import aliases from simple_data_loader import SimpleLeaderboardViewer from leaderboard_transformer import ( DataTransformer, transform_raw_dataframe, create_pretty_tag_map, INFORMAL_TO_FORMAL_NAME_MAP, _plot_scatter_plotly, format_cost_column, format_score_column, get_pareto_df, clean_llm_base_list, ) from config import ( CONFIG_NAME, EXTRACTED_DATA_DIR, IS_INTERNAL, RESULTS_DATASET, ) from content import ( create_gradio_anchor_id, format_error, get_benchmark_description, hf_uri_to_web_url, hyperlink, SCATTER_DISCLAIMER, ) api = HfApi() os.makedirs(EXTRACTED_DATA_DIR, exist_ok=True) # Company logo mapping - maps model name patterns to company logo files COMPANY_LOGO_MAP = { "anthropic": {"path": "assets/logo-anthropic.svg", "name": "Anthropic"}, "claude": {"path": "assets/logo-anthropic.svg", "name": "Anthropic"}, "openai": {"path": "assets/logo-openai.svg", "name": "OpenAI"}, "gpt": {"path": "assets/logo-openai.svg", "name": "OpenAI"}, "o1": {"path": "assets/logo-openai.svg", "name": "OpenAI"}, "o3": {"path": "assets/logo-openai.svg", "name": "OpenAI"}, "google": {"path": "assets/logo-google.svg", "name": "Google"}, "gemini": {"path": "assets/logo-google.svg", "name": "Google"}, "gemma": {"path": "assets/logo-google.svg", "name": "Google"}, "meta": {"path": "assets/logo-meta.svg", "name": "Meta"}, "llama": {"path": "assets/logo-meta.svg", "name": "Meta"}, "mistral": {"path": "assets/logo-mistral.svg", "name": "Mistral"}, "mixtral": {"path": "assets/logo-mistral.svg", "name": "Mistral"}, "codestral": {"path": "assets/logo-mistral.svg", "name": "Mistral"}, "deepseek": {"path": "assets/logo-deepseek.svg", "name": "DeepSeek"}, "xai": {"path": "assets/logo-xai.svg", "name": "xAI"}, "grok": {"path": "assets/logo-xai.svg", "name": "xAI"}, "cohere": {"path": "assets/logo-cohere.svg", "name": "Cohere"}, "command": {"path": "assets/logo-cohere.svg", "name": "Cohere"}, "qwen": {"path": "assets/logo-qwen.svg", "name": "Qwen"}, "alibaba": {"path": "assets/logo-qwen.svg", "name": "Qwen"}, "kimi": {"path": "assets/logo-moonshot.svg", "name": "Moonshot"}, "moonshot": {"path": "assets/logo-moonshot.svg", "name": "Moonshot"}, "minimax": {"path": "assets/logo-minimax.svg", "name": "MiniMax"}, } def get_company_from_model(model_name: str) -> dict: """ Gets the company info (logo path and name) from a model name. Returns default unknown logo if no match found. """ if not model_name: return {"path": "assets/logo-unknown.svg", "name": "Unknown"} # Handle list of models - use the first one if isinstance(model_name, list): model_name = model_name[0] if model_name else "" model_lower = str(model_name).lower() # Check each pattern for pattern, company_info in COMPANY_LOGO_MAP.items(): if pattern in model_lower: return company_info return {"path": "assets/logo-unknown.svg", "name": "Unknown"} def get_company_logo_html(model_name: str) -> str: """ Generates HTML for a company logo based on the model name. """ company_info = get_company_from_model(model_name) uri = get_svg_as_data_uri(company_info["path"]) if uri: return f' ${company_info[$ ' return "" # Simplified icon map (no tooling distinction, only openness) # Not actually used since we removed icons from the table, but keeping for potential future use OPENNESS_ICON_MAP = { aliases.CANONICAL_OPENNESS_OPEN: "assets/ellipse-pink.svg", aliases.CANONICAL_OPENNESS_CLOSED: "assets/ellipse-yellow.svg", } # Add aliases for canonical_openness, openness_aliases in aliases.OPENNESS_ALIASES.items(): for openness_alias in openness_aliases: OPENNESS_ICON_MAP[openness_alias] = OPENNESS_ICON_MAP[canonical_openness] OPENNESS_SVG_MAP = { aliases.CANONICAL_OPENNESS_OPEN: { "path": "assets/ellipse-pink.svg", "description": "Open source model" }, aliases.CANONICAL_OPENNESS_CLOSED: { "path": "assets/ellipse-yellow.svg", "description": "Closed source model" }, } def get_svg_as_data_uri(path: str) -> str: """Reads an SVG file and returns it as a base64-encoded data URI.""" try: with open(path, "rb") as svg_file: encoded_svg = base64.b64encode(svg_file.read()).decode("utf-8") return f"data:image/svg+xml;base64,{encoded_svg}" except FileNotFoundError: print(f"Warning: SVG file not found at {path}") return "" def create_svg_html(value, svg_map): """ Generates the absolute simplest HTML for an icon, without any extra text. This version is compatible with gr.DataFrame. """ if pd.isna(value) or value not in svg_map: return "" path_info = svg_map[value] # Handle both old string format and new object format if isinstance(path_info, dict): path = path_info["path"] else: path = path_info src = get_svg_as_data_uri(path) # Generate the HTML for the single icon, with NO text. if src: return f' {value}

' return "" def build_openness_tooltip_content() -> str: """ Generates the inner HTML for the Model Openness tooltip card using custom SVG lock icons. """ open_uri = get_svg_as_data_uri("assets/lock-open.svg") closed_uri = get_svg_as_data_uri("assets/lock-closed.svg") html_items = [ f""" """, f""" """ ] joined_items = "".join(html_items) return f"""""" def build_pareto_tooltip_content() -> str: """Generates the inner HTML for the Pareto tooltip card with final copy.""" trophy_uri = get_svg_as_data_uri("assets/trophy.svg") trophy_icon_html = f'

' return f"""

On Pareto Frontier

""" def build_descriptions_tooltip_content(table) -> str: """Generates the inner HTML for the Column Descriptions tooltip card depending on which kind of table.""" if table == "Overall": return """ """ elif table in ["Issue Resolution", "Frontend", "Greenfield", "Testing", "Information Gathering"]: return f""" """ else: # Fallback for any other table type, e.g., individual benchmarks return f""" """ # Create HTML for the "Openness" legend items for table using custom SVG lock icons open_lock_uri = get_svg_as_data_uri("assets/lock-open.svg") closed_lock_uri = get_svg_as_data_uri("assets/lock-closed.svg") openness_html_items = [ f'

' f'

' f'Open' f'

', f'

' f'

' f'Closed' f'

' ] openness_html = " ".join(openness_html_items) pareto_tooltip_content = build_pareto_tooltip_content() openness_tooltip_content = build_openness_tooltip_content() def create_legend_markdown(which_table: str) -> str: """ Generates the complete HTML for the legend section, including tooltips. This is used in the main leaderboard display. """ descriptions_tooltip_content = build_descriptions_tooltip_content(which_table) trophy_uri = get_svg_as_data_uri("assets/trophy.svg") # Add download section for benchmark-specific tables (not Overall or category pages) download_section = "" if which_table not in ["Overall", "Issue Resolution", "Frontend", "Greenfield", "Testing", "Information Gathering"]: download_section = """

Download

⬇️ Trajectories

""" legend_markdown = f"""

Pareto

On frontier

Model Openness {openness_tooltip_content}

{openness_html}

{download_section}

Column Descriptions

""" return legend_markdown # Create HTML for plot legend with company logos company_legend_items = [] # Show a sample of company logos in the legend sample_companies = [ ("Anthropic", "assets/logo-anthropic.svg"), ("OpenAI", "assets/logo-openai.svg"), ("Google", "assets/logo-google.svg"), ("Meta", "assets/logo-meta.svg"), ("Mistral", "assets/logo-mistral.svg"), ] for name, path in sample_companies: uri = get_svg_as_data_uri(path) if uri: company_legend_items.append( f'

' f' {name}

' f'{name}' f'

' ) plot_legend_html = f"""

Pareto

On frontier

Company Logos

{''.join(company_legend_items)}

"""; # --- Global State for Viewers (simple caching with TTL) --- CACHED_VIEWERS = {} CACHED_TAG_MAPS = {} _cache_lock = __import__('threading').Lock() _data_version = 0 # Incremented when data is refreshed def get_data_version(): """Get the current data version number.""" global _data_version return _data_version def clear_viewer_cache(): """ Clear all cached viewers and tag maps. Called when data is refreshed from the background scheduler. """ global CACHED_VIEWERS, CACHED_TAG_MAPS, _data_version with _cache_lock: CACHED_VIEWERS.clear() CACHED_TAG_MAPS.clear() _data_version += 1 print(f"[CACHE] Viewer cache cleared after data refresh (version: {_data_version})") # Register the cache clear callback with the data refresh system try: from setup_data import register_refresh_callback register_refresh_callback(clear_viewer_cache) except ImportError: pass # setup_data may not be available during import class DummyViewer: """A mock viewer to be cached on error. It has a ._load() method to ensure it behaves like the real LeaderboardViewer.""" def __init__(self, error_df): self._error_df = error_df def _load(self): # The _load method returns the error DataFrame and an empty tag map return self._error_df, {} def get_leaderboard_viewer_instance(split: str): """ Fetches the LeaderboardViewer for a split, using a thread-safe cache to avoid re-downloading data. On error, returns a stable DummyViewer object. """ global CACHED_VIEWERS, CACHED_TAG_MAPS with _cache_lock: if split in CACHED_VIEWERS: # Cache hit: return the cached viewer and tag map return CACHED_VIEWERS[split], CACHED_TAG_MAPS.get(split, {"Overall": []}) # --- Cache miss: try to load data from the source --- try: # First try to load from extracted data directory (local mock data) data_dir = EXTRACTED_DATA_DIR if os.path.exists(EXTRACTED_DATA_DIR) else "mock_results" print(f"Loading data for split '{split}' from: {data_dir}/{CONFIG_NAME}") viewer = SimpleLeaderboardViewer( data_dir=data_dir, config=CONFIG_NAME, split=split ) # Simplify tag map creation pretty_tag_map = create_pretty_tag_map(viewer.tag_map, INFORMAL_TO_FORMAL_NAME_MAP) # Cache the results for next time (thread-safe) with _cache_lock: CACHED_VIEWERS[split] = viewer CACHED_TAG_MAPS[split] = pretty_tag_map # Cache the pretty map directly return viewer, pretty_tag_map except Exception as e: # On ANY error, create a consistent error message and cache a DummyViewer error_message = f"Error loading data for split '{split}': {e}" print(format_error(error_message)) dummy_df = pd.DataFrame({"Message": [error_message]}) dummy_viewer = DummyViewer(dummy_df) dummy_tag_map = {"Overall": []} # Cache the dummy objects so we don't try to fetch again on this run with _cache_lock: CACHED_VIEWERS[split] = dummy_viewer CACHED_TAG_MAPS[split] = dummy_tag_map return dummy_viewer, dummy_tag_map def create_leaderboard_display( full_df: pd.DataFrame, tag_map: dict, category_name: str, split_name: str ): """ This UI factory takes pre-loaded data and renders the main DataFrame and Plot for a given category (e.g., "Overall" or "Literature Understanding"). The display includes a timer that periodically checks for data updates and refreshes the UI when new data is available. """ # Track the data version when this display was created initial_data_version = get_data_version() # 1. Instantiate the transformer and get the specific view for this category. # The function no longer loads data itself; it filters the data it receives. transformer = DataTransformer(full_df, tag_map) df_view_full, plots_dict = transformer.view(tag=category_name, use_plotly=True) def prepare_df_for_display(df_view): """Prepare a DataFrame for display with all formatting applied.""" df_display = df_view.copy() # Get Pareto frontier info pareto_df = get_pareto_df(df_display) trophy_uri = get_svg_as_data_uri("assets/trophy.svg") if not pareto_df.empty and 'id' in pareto_df.columns: pareto_agent_names = pareto_df['id'].tolist() else: pareto_agent_names = [] for col in df_display.columns: if "Cost" in col: df_display = format_cost_column(df_display, col) for col in df_display.columns: if "Score" in col: df_display = format_score_column(df_display, col) # Clean the Language Model column first df_display['Language Model'] = df_display['Language Model'].apply(clean_llm_base_list) # Now combine icons with Language Model column def format_language_model_with_icons(row): icons_html = '' # Add Pareto trophy if on frontier if row['id'] in pareto_agent_names: icons_html += f' On Pareto Frontier

' # Add openness lock icon openness_val = row.get('Openness', '') if openness_val in [aliases.CANONICAL_OPENNESS_OPEN, 'Open', 'Open Source', 'Open Source + Open Weights']: lock_uri = get_svg_as_data_uri("assets/lock-open.svg") icons_html += f' Open

' else: lock_uri = get_svg_as_data_uri("assets/lock-closed.svg") icons_html += f' Closed

' # Add company logo company_html = get_company_logo_html(row['Language Model']) if company_html: icons_html += company_html # Format the model name model_name = row['Language Model'] if isinstance(model_name, list): if len(model_name) > 1: tooltip_text = "\\n".join(map(str, model_name)) model_text = f'' elif len(model_name) == 1: model_text = model_name[0] else: model_text = str(model_name) else: model_text = str(model_name) # Wrap in a flex container to keep icons horizontal return f'

{icons_html}{model_text}

' df_display['Language Model'] = df_display.apply(format_language_model_with_icons, axis=1) if 'Source' in df_display.columns: df_display['SDK Version'] = df_display.apply( lambda row: f"{row['SDK Version']} {row['Source']}" if pd.notna(row['Source']) and row['Source'] else row['SDK Version'], axis=1 ) columns_to_drop = ['id', 'Openness', 'Agent Tooling', 'Source'] df_display = df_display.drop(columns=columns_to_drop, errors='ignore') return df_display # Prepare both complete and all entries versions # Complete entries have all 5 categories submitted # The 'Categories Attempted' column is formatted as "X/5" if 'Categories Attempted' in df_view_full.columns: df_view_complete = df_view_full[df_view_full['Categories Attempted'] == '5/5'].copy() else: df_view_complete = df_view_full.copy() df_display_complete = prepare_df_for_display(df_view_complete) df_display_all = prepare_df_for_display(df_view_full) # If no complete entries exist, show all entries by default has_complete_entries = len(df_display_complete) > 0 # Determine primary score/cost columns for scatter plot if category_name == "Overall": primary_score_col = "Average Score" primary_cost_col = "Average Cost" else: primary_score_col = f"{category_name} Score" primary_cost_col = f"{category_name} Cost" # Function to create scatter plot from data def create_scatter_plot(df_data): return _plot_scatter_plotly( data=df_data, x=primary_cost_col if primary_cost_col in df_data.columns else None, y=primary_score_col if primary_score_col in df_data.columns else "Average Score", agent_col="SDK Version", name=category_name ) # Create initial scatter plots for both complete and all data scatter_plot_complete = create_scatter_plot(df_view_complete) if has_complete_entries else go.Figure() scatter_plot_all = create_scatter_plot(df_view_full) # Now get headers from the renamed dataframe (use all entries to ensure headers are present) df_headers = df_display_all.columns.tolist() df_datatypes = [] for col in df_headers: if col == "Logs" or "Cost" in col or "Score" in col: df_datatypes.append("markdown") elif col in ["SDK Version", "Language Model"]: df_datatypes.append("html") else: df_datatypes.append("str") # Dynamically set widths for the DataFrame columns # Order: Language Model, SDK Version, Average Score, Average Cost, ... fixed_start_widths = [280, 100, 100] # Language Model (with icons), SDK Version, Average Score num_score_cost_cols = 0 remaining_headers = df_headers[len(fixed_start_widths):] for col in remaining_headers: if "Score" in col or "Cost" in col: num_score_cost_cols += 1 dynamic_widths = [90] * num_score_cost_cols fixed_end_widths = [90, 100, 50] # Categories Attempted, Date, Logs # 5. Combine all the lists to create the final, fully dynamic list. final_column_widths = fixed_start_widths + dynamic_widths + fixed_end_widths # Calculate counts for the checkbox label num_complete = len(df_display_complete) num_total = len(df_display_all) num_incomplete = num_total - num_complete # Add toggle for showing incomplete entries ABOVE the plot if has_complete_entries: show_incomplete_checkbox = gr.Checkbox( label=f"Show incomplete entries ({num_incomplete} entries with fewer than 5 categories)", value=False, elem_id="show-incomplete-toggle" ) else: show_incomplete_checkbox = None gr.Markdown(f"*No entries with all 5 categories completed yet. Showing all {num_total} entries.*") # Plot component - show complete entries by default if available initial_plot = scatter_plot_complete if has_complete_entries else scatter_plot_all plot_component = gr.Plot( value=initial_plot, show_label=False, ) gr.Markdown(value=SCATTER_DISCLAIMER, elem_id="scatter-disclaimer") # Put table and key into an accordion with gr.Accordion("Show / Hide Table View", open=True, elem_id="leaderboard-accordion"): # If there are complete entries, show toggle. If not, show all entries. if has_complete_entries: # Start with complete entries only (default) dataframe_component = gr.DataFrame( headers=df_headers, value=df_display_complete, datatype=df_datatypes, interactive=False, wrap=True, column_widths=final_column_widths, elem_classes=["wrap-header-df"], show_search="search", elem_id="main-leaderboard" ) # Update function for the toggle - updates both table and plot def update_display(show_incomplete): if show_incomplete: return df_display_all, scatter_plot_all else: return df_display_complete, scatter_plot_complete show_incomplete_checkbox.change( fn=update_display, inputs=[show_incomplete_checkbox], outputs=[dataframe_component, plot_component] ) else: dataframe_component = gr.DataFrame( headers=df_headers, value=df_display_all, datatype=df_datatypes, interactive=False, wrap=True, column_widths=final_column_widths, elem_classes=["wrap-header-df"], show_search="search", elem_id="main-leaderboard" ) legend_markdown = create_legend_markdown(category_name) gr.HTML(value=legend_markdown, elem_id="legend-markdown") # Add a timer to periodically check for data updates and refresh the UI # This runs every 60 seconds to check if new data is available def check_and_refresh_data(current_checkbox_state): """Check if data has been refreshed and return updated data if so.""" current_version = get_data_version() if current_version > initial_data_version: # Data has been refreshed, reload it print(f"[REFRESH] Data version changed from {initial_data_version} to {current_version}, reloading...") new_df, new_tag_map = get_full_leaderboard_data(split_name) if not new_df.empty: new_transformer = DataTransformer(new_df, new_tag_map) new_df_view_full, _ = new_transformer.view(tag=category_name, use_plotly=True) # Prepare both complete and all entries versions if 'Categories Attempted' in new_df_view_full.columns: new_df_view_complete = new_df_view_full[new_df_view_full['Categories Attempted'] == '5/5'].copy() else: new_df_view_complete = new_df_view_full.copy() new_df_display_complete = prepare_df_for_display(new_df_view_complete) new_df_display_all = prepare_df_for_display(new_df_view_full) # Create new scatter plots new_scatter_complete = create_scatter_plot(new_df_view_complete) if len(new_df_display_complete) > 0 else go.Figure() new_scatter_all = create_scatter_plot(new_df_view_full) # Return the appropriate data based on checkbox state if current_checkbox_state: return new_df_display_all, new_scatter_all else: return new_df_display_complete, new_scatter_complete # No change, return current values if current_checkbox_state: return df_display_all, scatter_plot_all else: return df_display_complete, scatter_plot_complete # Create a timer that checks for updates every 60 seconds refresh_timer = gr.Timer(value=60) # Connect the timer to the refresh function if show_incomplete_checkbox is not None: refresh_timer.tick( fn=check_and_refresh_data, inputs=[show_incomplete_checkbox], outputs=[dataframe_component, plot_component] ) else: # If no checkbox, always show all data def check_and_refresh_all(): current_version = get_data_version() if current_version > initial_data_version: print(f"[REFRESH] Data version changed, reloading...") new_df, new_tag_map = get_full_leaderboard_data(split_name) if not new_df.empty: new_transformer = DataTransformer(new_df, new_tag_map) new_df_view_full, _ = new_transformer.view(tag=category_name, use_plotly=True) new_df_display_all = prepare_df_for_display(new_df_view_full) new_scatter_all = create_scatter_plot(new_df_view_full) return new_df_display_all, new_scatter_all return df_display_all, scatter_plot_all refresh_timer.tick( fn=check_and_refresh_all, inputs=[], outputs=[dataframe_component, plot_component] ) # Return the components so they can be referenced elsewhere. return plot_component, dataframe_component # # --- Detailed Benchmark Display --- def create_benchmark_details_display( full_df: pd.DataFrame, tag_map: dict, category_name: str, validation: bool = False, ): """ Generates a detailed breakdown for each benchmark within a given category. For each benchmark, it creates a title, a filtered table, and a scatter plot. Args: full_df (pd.DataFrame): The complete, "pretty" dataframe for the entire split. tag_map (dict): The "pretty" tag map to find the list of benchmarks. category_name (str): The main category to display details for (e.g., "Literature Understanding"). """ # 1. Get the list of benchmarks for the selected category benchmark_names = tag_map.get(category_name, []) if not benchmark_names: gr.Markdown(f"No detailed benchmarks found for the category: {category_name}") return gr.HTML(f'

{category_name} Detailed Benchmark Results

') gr.Markdown("---") # 2. Loop through each benchmark and create its UI components for benchmark_name in benchmark_names: anchor_id = create_gradio_anchor_id(benchmark_name, validation) gr.HTML( f"""

{benchmark_name} Leaderboard 🔗

{get_benchmark_description(benchmark_name, validation)}

""" ) # 3. Prepare the data for this specific benchmark's table and plot benchmark_score_col = f"{benchmark_name} Score" benchmark_cost_col = f"{benchmark_name} Cost" benchmark_download_col = f"{benchmark_name} Download" # Define the columns needed for the detailed table table_cols = ['SDK Version','Source','Openness', 'Date', benchmark_score_col, benchmark_cost_col,'Logs', benchmark_download_col, 'id', 'Language Model'] # Filter to only columns that actually exist in the full dataframe existing_table_cols = [col for col in table_cols if col in full_df.columns] if benchmark_score_col not in existing_table_cols: gr.Markdown(f"Score data for {benchmark_name} not available.") continue # Skip to the next benchmark if score is missing # Create a specific DataFrame for the table view benchmark_table_df = full_df[existing_table_cols].copy() pareto_df = get_pareto_df(benchmark_table_df) # Get the list of agents on the frontier. We'll use this list later. trophy_uri = get_svg_as_data_uri("assets/trophy.svg") if not pareto_df.empty and 'id' in pareto_df.columns: pareto_agent_names = pareto_df['id'].tolist() else: pareto_agent_names = [] # Clean the Language Model column first benchmark_table_df['Language Model'] = benchmark_table_df['Language Model'].apply(clean_llm_base_list) # Combine icons with Language Model column def format_language_model_with_icons(row): icons_html = '' # Add Pareto trophy if on frontier if row['id'] in pareto_agent_names: icons_html += f' On Pareto Frontier

' else: lock_uri = get_svg_as_data_uri("assets/lock-closed.svg") icons_html += f' Closed

{icons_html}{model_text}

' benchmark_table_df['Language Model'] = benchmark_table_df.apply(format_language_model_with_icons, axis=1) # append the repro url to the end of the SDK Version if 'Source' in benchmark_table_df.columns: benchmark_table_df['SDK Version'] = benchmark_table_df.apply( lambda row: f"{row['SDK Version']} {row['Source']}" if row['Source'] else row['SDK Version'], axis=1 ) # Calculated and add "Benchmark Attempted" column def check_benchmark_status(row): has_score = pd.notna(row.get(benchmark_score_col)) has_cost = pd.notna(row.get(benchmark_cost_col)) if has_score and has_cost: return "✅" if has_score or has_cost: return "⚠️" return "🚫 " # Apply the function to create the new column benchmark_table_df['Attempted Benchmark'] = benchmark_table_df.apply(check_benchmark_status, axis=1) # Sort the DataFrame if benchmark_score_col in benchmark_table_df.columns: benchmark_table_df = benchmark_table_df.sort_values( by=benchmark_score_col, ascending=False, na_position='last' ) # 1. Format the cost and score columns benchmark_table_df = format_cost_column(benchmark_table_df, benchmark_cost_col) benchmark_table_df = format_score_column(benchmark_table_df, benchmark_score_col) # Format download column as clickable icon if benchmark_download_col in benchmark_table_df.columns: def format_download_link(url): if pd.isna(url) or url == "": return "" return f"[⬇️]({url})" benchmark_table_df[benchmark_download_col] = benchmark_table_df[benchmark_download_col].apply(format_download_link) desired_cols_in_order = [ 'Language Model', 'SDK Version', 'Attempted Benchmark', benchmark_score_col, benchmark_cost_col, 'Date', 'Logs', benchmark_download_col ] for col in desired_cols_in_order: if col not in benchmark_table_df.columns: benchmark_table_df[col] = pd.NA # Add as an empty column benchmark_table_df = benchmark_table_df[desired_cols_in_order] # Rename columns for a cleaner table display, as requested benchmark_table_df.rename(columns={ benchmark_score_col: 'Score', benchmark_cost_col: 'Cost', benchmark_download_col: '⬇️', # Empty-ish header with icon hint }, inplace=True) # Now get headers from the renamed dataframe df_headers = benchmark_table_df.columns.tolist() df_datatypes = [] for col in df_headers: if col in ["Logs", "⬇️"] or "Cost" in col or "Score" in col: df_datatypes.append("markdown") elif col in ["SDK Version", "Language Model"]: df_datatypes.append("html") else: df_datatypes.append("str") benchmark_plot = _plot_scatter_plotly( data=full_df, x=benchmark_cost_col, y=benchmark_score_col, agent_col="SDK Version", name=benchmark_name ) gr.Plot(value=benchmark_plot, show_label=False) gr.Markdown(value=SCATTER_DISCLAIMER, elem_id="scatter-disclaimer") # Put table and key into an accordion with gr.Accordion("Show / Hide Table View", open=True, elem_id="leaderboard-accordion"): gr.DataFrame( headers=df_headers, value=benchmark_table_df, datatype=df_datatypes, interactive=False, wrap=True, column_widths=[200, 80, 40, 80, 80, 150, 40, 40], # Language Model, SDK Version, Attempted, Score, Cost, Date, Logs, Download show_search="search", elem_classes=["wrap-header-df"] ) legend_markdown = create_legend_markdown(benchmark_name) gr.HTML(value=legend_markdown, elem_id="legend-markdown") def get_full_leaderboard_data(split: str) -> tuple[pd.DataFrame, dict]: """ Loads and transforms the complete dataset for a given split. This function handles caching and returns the final "pretty" DataFrame and tag map. """ viewer_or_data, raw_tag_map = get_leaderboard_viewer_instance(split) if isinstance(viewer_or_data, (SimpleLeaderboardViewer, DummyViewer)): raw_df, _ = viewer_or_data._load() if raw_df.empty: return pd.DataFrame(), {} pretty_df = transform_raw_dataframe(raw_df) pretty_tag_map = create_pretty_tag_map(raw_tag_map, INFORMAL_TO_FORMAL_NAME_MAP) if "Logs" in pretty_df.columns: def format_log_entry_to_html(raw_uri): if pd.isna(raw_uri) or raw_uri == "": return "" web_url = hf_uri_to_web_url(str(raw_uri)) return hyperlink(web_url, "🔗") if web_url else "" # Apply the function to the "Logs" column pretty_df["Logs"] = pretty_df["Logs"].apply(format_log_entry_to_html) if "Source" in pretty_df.columns: def format_source_url_to_html(raw_url): # Handle empty or NaN values, returning a blank string. if pd.isna(raw_url) or raw_url == "": return "" # Assume 'source_url' is already a valid web URL and doesn't need conversion. return hyperlink(str(raw_url), "🔗") # Apply the function to the "source_url" column. pretty_df["Source"] = pretty_df["Source"].apply(format_source_url_to_html) return pretty_df, pretty_tag_map # Fallback for unexpected types return pd.DataFrame(), {} def create_sub_navigation_bar(tag_map: dict, category_name: str, validation: bool = False) -> gr.HTML: """ Builds the entire sub-navigation bar as a single, self-contained HTML component. This bypasses Gradio's layout components, giving us full control. """ benchmark_names = tag_map.get(category_name, []) if not benchmark_names: # Return an empty HTML component to prevent errors return gr.HTML() # Start building the list of HTML button elements as strings html_buttons = [] for name in benchmark_names: target_id = create_gradio_anchor_id(name, validation) # Create a standard HTML button. # The onclick attribute calls our global JS function directly. # Note the mix of double and single quotes. button_str = f""" """ html_buttons.append(button_str) # Join the button strings and wrap them in a single div container # This container will be our flexbox row. full_html = f""" """ # Return the entire navigation bar as one single Gradio HTML component return gr.HTML(full_html) def format_llm_base_with_html(value): """ Formats the 'Models Used' cell value. If the value is a list with more than 1 element, it returns an HTML with the full list in a hover-over tooltip. If it's a single-element list, it returns just that element. Otherwise, it returns the original value. """ if isinstance(value, list): if len(value) > 1: # Join the list items with a newline character for a clean tooltip tooltip_text = "\n".join(map(str, value)) # Return an HTML span with the title attribute for the tooltip return f'' if len(value) == 1: # If only one item, just return that item return value[0] # Return the value as-is if it's not a list or is an empty list return value