import os # Set HF_HOME for caching os.environ["HF_HOME"] = "src/data_cache" import streamlit as st import pandas as pd import altair as alt from huggingface_hub import HfApi, HfFileSystem import json from pathlib import Path from typing import Dict, List, Optional import numpy as np # Page config st.set_page_config( page_title="Grounding Benchmark Leaderboard", page_icon="🎯", layout="wide" ) # Constants REPO_ID = "mlfoundations-cua-dev/leaderboard" GROUNDING_PATH = "grounding" # Baselines for different datasets BASELINES = { "screenspot-v2": { "Qwen2-VL-7B": { "desktop_text": 52.01, "desktop_icon": 44.98, "web_text": 33.04, "web_icon": 21.84, "overall": 37.96 }, "UI-TARS-2B": { "desktop_text": 90.7, "desktop_icon": 68.6, "web_text": 87.2, "web_icon": 84.7, "overall": 82.8 }, "UI-TARS-7B": { "desktop_text": 95.4, "desktop_icon": 87.8, "web_text": 93.8, "web_icon": 91.6, "overall": 92.2 }, "UI-TARS-72B": { "desktop_text": 91.2, "desktop_icon": 87.8, "web_text": 87.7, "web_icon": 86.3, "overall": 88.3 }, "Qwen2.5-VL-3B-Instruct": {"desktop_text": 54.1, "desktop_icon": 30.0, "web_text": 31.2, "web_icon": 48.3, "overall": 41.5}, "Qwen2.5-VL-7B-Instruct": {"desktop_text": 87.6, "desktop_icon": 65.7, "web_text": 90.2, "web_icon": 79.8, "overall": 81.9}, }, "screenspot-pro": { "Qwen2.5-VL-3B-Instruct": {"overall": 16.1}, "Qwen2.5-VL-7B-Instruct": {"overall": 26.8}, "Qwen2.5-VL-72B-Instruct": {"overall": 53.3}, "UI-TARS-2B": {"overall": 27.7}, "UI-TARS-7B": {"overall": 35.7}, "UI-TARS-72B": {"overall": 38.1} }, "showdown-clicks": { "UI-TARS-2B": {"overall": 59.8}, "UI-TARS-7B": {"overall": 66.1}, "UI-TARS-1.5-7B": {"overall": 67.2}, } } @st.cache_data() # Cache without TTL - manual refresh only def fetch_leaderboard_data(): """Fetch all grounding results from HuggingFace leaderboard by streaming JSON files.""" api = HfApi() fs = HfFileSystem() try: # List all files in the grounding directory files = api.list_repo_files(repo_id=REPO_ID, repo_type="dataset") grounding_files = [f for f in files if f.startswith(f"{GROUNDING_PATH}/") and f.endswith(".json")] results = [] # Create progress bar for loading progress_bar = st.progress(0) status_text = st.empty() for idx, file_path in enumerate(grounding_files): try: # Update progress progress = (idx + 1) / len(grounding_files) progress_bar.progress(progress) status_text.text(f"Loading {idx + 1}/{len(grounding_files)} files...") # Stream the JSON file content directly from HuggingFace file_url = f"datasets/{REPO_ID}/{file_path}" # Read the file content directly without downloading with fs.open(file_url, 'r') as f: data = json.load(f) # Extract only the necessary information metadata = data.get("metadata", {}) metrics = data.get("metrics", {}) detailed_results = data.get("detailed_results", {}) # Parse the file path to get dataset and model info path_parts = file_path.split('/') dataset_name = path_parts[1] if len(path_parts) > 1 else "unknown" # Get model name from metadata or path model_checkpoint = metadata.get("model_checkpoint", "") model_name = model_checkpoint.split('/')[-1] base_model_name = None is_checkpoint = False # Handle checkpoint names if not model_name and len(path_parts) > 2: # Check if it's a checkpoint subdirectory structure if len(path_parts) > 3 and path_parts[2] != path_parts[3]: # Format: grounding/dataset/base_model/checkpoint.json base_model_name = path_parts[2] checkpoint_file = path_parts[3].replace(".json", "") model_name = f"{base_model_name}/{checkpoint_file}" is_checkpoint = True else: # Regular format: grounding/dataset/results_modelname.json model_name = path_parts[2].replace("results_", "").replace(".json", "") base_model_name = model_name # Check if model name indicates a checkpoint if 'checkpoint-' in model_name: is_checkpoint = True if not base_model_name: # Extract base model name from full path if '/' in model_name: parts = model_name.split('/') base_model_name = parts[0] else: # Try to get from model_checkpoint path checkpoint_parts = model_checkpoint.split('/') if len(checkpoint_parts) > 1: base_model_name = checkpoint_parts[-2] # Extract UI type results if available ui_type_results = detailed_results.get("by_ui_type", {}) dataset_type_results = detailed_results.get("by_dataset_type", {}) results_by_file = detailed_results.get("by_file", {}) # Create a compact result entry (only keep what we need for visualization) result_entry = { "dataset": dataset_name, "model": model_name, "base_model": base_model_name or model_name, "is_checkpoint": is_checkpoint, "model_path": model_checkpoint, "overall_accuracy": metrics.get("accuracy", 0) * 100, # Convert to percentage "total_samples": metrics.get("total", 0), "timestamp": metadata.get("evaluation_timestamp", ""), "checkpoint_steps": metadata.get("checkpoint_steps"), "training_loss": metadata.get("training_loss"), "ui_type_results": ui_type_results, "dataset_type_results": dataset_type_results, "results_by_file": results_by_file } results.append(result_entry) except Exception as e: st.warning(f"Error loading {file_path}: {str(e)}") continue # Clear progress indicators progress_bar.empty() status_text.empty() # Create DataFrame df = pd.DataFrame(results) # Process checkpoints: for each base model, find the best checkpoint if not df.empty: # Group by dataset and base_model grouped = df.groupby(['dataset', 'base_model']) # For each group, find the best checkpoint best_models = [] for (dataset, base_model), group in grouped: if len(group) > 1: # Multiple entries for this model (likely checkpoints) best_idx = group['overall_accuracy'].idxmax() best_row = group.loc[best_idx].copy() # Check if the best is the last checkpoint checkpoint_steps = group[group['checkpoint_steps'].notna()]['checkpoint_steps'].sort_values() if len(checkpoint_steps) > 0: last_checkpoint_steps = checkpoint_steps.iloc[-1] best_checkpoint_steps = best_row['checkpoint_steps'] if pd.notna(best_checkpoint_steps) and best_checkpoint_steps != last_checkpoint_steps: # Best checkpoint is not the last one, add asterisk best_row['model'] = best_row['model'] + '*' best_row['is_best_not_last'] = True else: best_row['is_best_not_last'] = False # Store all checkpoints for this model best_row['all_checkpoints'] = group.to_dict('records') best_models.append(best_row) else: # Single entry for this model row = group.iloc[0].copy() row['is_best_not_last'] = False row['all_checkpoints'] = [row.to_dict()] best_models.append(row) # Create new dataframe with best models df_best = pd.DataFrame(best_models) return df_best return df except Exception as e: st.error(f"Error fetching leaderboard data: {str(e)}") return pd.DataFrame() def parse_ui_type_metrics(df: pd.DataFrame, dataset_filter: str) -> pd.DataFrame: """Parse UI type metrics from the results dataframe.""" metrics_list = [] for _, row in df.iterrows(): if row['dataset'] != dataset_filter: continue model = row['model'] ui_results = row.get('ui_type_results', {}) dataset_type_results = row.get('dataset_type_results', {}) results_by_file = row.get('results_by_file', {}) # For ScreenSpot datasets if 'screenspot' in dataset_filter.lower(): # Check if we have desktop/web breakdown in results_by_file desktop_file = None web_file = None for filename, file_results in results_by_file.items(): if 'desktop' in filename.lower(): desktop_file = file_results elif 'web' in filename.lower(): web_file = file_results if desktop_file and web_file: # We have desktop/web breakdown desktop_text = desktop_file.get('by_ui_type', {}).get('text', {}).get('correct', 0) / max(desktop_file.get('by_ui_type', {}).get('text', {}).get('total', 1), 1) * 100 desktop_icon = desktop_file.get('by_ui_type', {}).get('icon', {}).get('correct', 0) / max(desktop_file.get('by_ui_type', {}).get('icon', {}).get('total', 1), 1) * 100 web_text = web_file.get('by_ui_type', {}).get('text', {}).get('correct', 0) / max(web_file.get('by_ui_type', {}).get('text', {}).get('total', 1), 1) * 100 web_icon = web_file.get('by_ui_type', {}).get('icon', {}).get('correct', 0) / max(web_file.get('by_ui_type', {}).get('icon', {}).get('total', 1), 1) * 100 # Calculate averages desktop_avg = (desktop_text + desktop_icon) / 2 if (desktop_text > 0 or desktop_icon > 0) else 0 web_avg = (web_text + web_icon) / 2 if (web_text > 0 or web_icon > 0) else 0 text_avg = (desktop_text + web_text) / 2 if (desktop_text > 0 or web_text > 0) else 0 icon_avg = (desktop_icon + web_icon) / 2 if (desktop_icon > 0 or web_icon > 0) else 0 # For screenspot-v2, calculate the overall as average of desktop and web if dataset_filter == 'screenspot-v2': overall = (desktop_avg + web_avg) / 2 if (desktop_avg > 0 or web_avg > 0) else row['overall_accuracy'] else: overall = row['overall_accuracy'] metrics_list.append({ 'model': model, 'desktop_text': desktop_text, 'desktop_icon': desktop_icon, 'web_text': web_text, 'web_icon': web_icon, 'desktop_avg': desktop_avg, 'web_avg': web_avg, 'text_avg': text_avg, 'icon_avg': icon_avg, 'overall': overall, 'is_best_not_last': row.get('is_best_not_last', False), 'all_checkpoints': row.get('all_checkpoints', []) }) elif 'text' in ui_results and 'icon' in ui_results: # Simple text/icon structure without desktop/web breakdown text_acc = (ui_results.get('text', {}).get('correct', 0) / max(ui_results.get('text', {}).get('total', 1), 1)) * 100 icon_acc = (ui_results.get('icon', {}).get('correct', 0) / max(ui_results.get('icon', {}).get('total', 1), 1)) * 100 metrics_list.append({ 'model': model, 'text': text_acc, 'icon': icon_acc, 'overall': row['overall_accuracy'], 'is_best_not_last': row.get('is_best_not_last', False), 'all_checkpoints': row.get('all_checkpoints', []) }) else: # Try to get from dataset_type_results if available found_data = False for dataset_key in dataset_type_results: if 'screenspot' in dataset_key.lower(): dataset_data = dataset_type_results[dataset_key] if 'by_ui_type' in dataset_data: ui_data = dataset_data['by_ui_type'] text_data = ui_data.get('text', {}) icon_data = ui_data.get('icon', {}) text_acc = (text_data.get('correct', 0) / max(text_data.get('total', 1), 1)) * 100 icon_acc = (icon_data.get('correct', 0) / max(icon_data.get('total', 1), 1)) * 100 metrics_list.append({ 'model': model, 'text': text_acc, 'icon': icon_acc, 'overall': row['overall_accuracy'], 'is_best_not_last': row.get('is_best_not_last', False), 'all_checkpoints': row.get('all_checkpoints', []) }) found_data = True break if not found_data: # No UI type data available, just use overall metrics_list.append({ 'model': model, 'overall': row['overall_accuracy'], 'is_best_not_last': row.get('is_best_not_last', False), 'all_checkpoints': row.get('all_checkpoints', []) }) else: # For non-screenspot datasets, just pass through overall accuracy metrics_list.append({ 'model': model, 'overall': row['overall_accuracy'], 'is_best_not_last': row.get('is_best_not_last', False), 'all_checkpoints': row.get('all_checkpoints', []) }) return pd.DataFrame(metrics_list) def create_bar_chart(data: pd.DataFrame, metric: str, title: str): """Create a bar chart for a specific metric.""" # Prepare data for the chart chart_data = [] # Add model results for _, row in data.iterrows(): if metric in row and row[metric] > 0: chart_data.append({ 'Model': row['model'], 'Score': row[metric], 'Type': 'Evaluated' }) # Add baselines if available dataset = st.session_state.get('selected_dataset', '') if dataset in BASELINES: for baseline_name, baseline_metrics in BASELINES[dataset].items(): metric_key = metric.replace('_avg', '').replace('avg', 'overall') if metric_key in baseline_metrics: chart_data.append({ 'Model': baseline_name, 'Score': baseline_metrics[metric_key], 'Type': 'Baseline' }) if not chart_data: return None df_chart = pd.DataFrame(chart_data) # Create the bar chart chart = alt.Chart(df_chart).mark_bar().encode( x=alt.X('Model:N', sort=alt.EncodingSortField(field='Score', order='descending'), axis=alt.Axis(labelAngle=-45)), y=alt.Y('Score:Q', scale=alt.Scale(domain=[0, 100]), axis=alt.Axis(title='Score (%)')), color=alt.Color('Type:N', scale=alt.Scale(domain=['Evaluated', 'Baseline'], range=['#4ECDC4', '#FFA726'])), tooltip=['Model', 'Score', 'Type'] ).properties( title=title, width=500, height=400 ) # Add value labels text = chart.mark_text( align='center', baseline='bottom', dy=-5 ).encode( text=alt.Text('Score:Q', format='.1f') ) return chart + text def main(): st.title("🎯 Grounding Benchmark Leaderboard") st.markdown("Visualization of model performance on grounding benchmarks") # Add refresh button col1, col2 = st.columns([1, 5]) with col1: if st.button("🔄 Refresh", help="Refresh leaderboard data from HuggingFace"): st.cache_data.clear() st.rerun() # Fetch data with st.spinner("Loading leaderboard data..."): df = fetch_leaderboard_data() if df.empty: st.warning("No data available in the leaderboard.") return # Sidebar filters st.sidebar.header("Filters") # Dataset filter datasets = sorted(df['dataset'].unique()) selected_dataset = st.sidebar.selectbox("Select Dataset", datasets) st.session_state['selected_dataset'] = selected_dataset # Filter data filtered_df = df[df['dataset'] == selected_dataset] # Model filter - changed to multiselect for selective visualization st.sidebar.subheader("Select Models to Display") all_models = sorted(filtered_df['model'].unique()) # Add "Select All" / "Deselect All" buttons col1, col2 = st.sidebar.columns(2) with col1: if st.button("Select All", key="select_all"): st.session_state['selected_models'] = all_models with col2: if st.button("Deselect All", key="deselect_all"): st.session_state['selected_models'] = [] # Initialize selected models if not in session state if 'selected_models' not in st.session_state: st.session_state['selected_models'] = all_models # Multi-select widget for models selected_models = st.sidebar.multiselect( "Models to visualize:", options=all_models, default=st.session_state.get('selected_models', all_models), key='model_multiselect' ) # Update session state st.session_state['selected_models'] = selected_models # Filter dataframe based on selected models if selected_models: filtered_df = filtered_df[filtered_df['model'].isin(selected_models)] else: # If no models selected, show empty dataframe filtered_df = pd.DataFrame() # Main content st.header(f"Results for {selected_dataset}") # Check if any models are selected if filtered_df.empty: st.warning("No models selected. Please select at least one model from the sidebar to visualize results.") return # Overall metrics col1, col2, col3 = st.columns(3) with col1: st.metric("Models Evaluated", len(filtered_df)) with col2: if not filtered_df.empty: best_acc = filtered_df['overall_accuracy'].max() best_model = filtered_df[filtered_df['overall_accuracy'] == best_acc]['model'].iloc[0] st.metric("Best Overall Accuracy", f"{best_acc:.1f}%", help=f"Model: {best_model}") with col3: total_samples = filtered_df['total_samples'].sum() st.metric("Total Samples Evaluated", f"{total_samples:,}") # Parse UI type metrics ui_metrics_df = parse_ui_type_metrics(filtered_df, selected_dataset) # Add metric selector for screenspot datasets selected_metric = 'overall' # Default metric if not ui_metrics_df.empty: # Metric selector dropdown if selected_dataset == 'screenspot-v2': metric_options = { 'overall': 'Overall Average (Desktop + Web) / 2', 'desktop_text': 'Desktop (Text)', 'desktop_icon': 'Desktop (Icon)', 'web_text': 'Web (Text)', 'web_icon': 'Web (Icon)', } else: # For screenspot-pro and showdown-clicks, only show overall average metric_options = { 'overall': 'Overall Average' } selected_metric = st.selectbox( "Select metric to visualize:", options=list(metric_options.keys()), format_func=lambda x: metric_options[x], key="metric_selector" ) # Add note about asterisks if any(ui_metrics_df['is_best_not_last']): st.info("* indicates the best checkpoint is not the last checkpoint") # Create single chart for selected metric chart = create_bar_chart(ui_metrics_df, selected_metric, metric_options[selected_metric]) if chart: st.altair_chart(chart, use_container_width=True) else: st.warning(f"No data available for {metric_options[selected_metric]}") if __name__ == "__main__": main()