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 } } } @st.cache_data(ttl=300) # Cache for 5 minutes 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] # 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 = path_parts[2] checkpoint_file = path_parts[3].replace(".json", "") model_name = f"{base_model}/{checkpoint_file}" else: # Regular format: grounding/dataset/results_modelname.json model_name = path_parts[2].replace("results_", "").replace(".json", "") # Extract UI type results if available ui_type_results = detailed_results.get("by_ui_type", {}) dataset_type_results = detailed_results.get("by_dataset_type", {}) # Create a compact result entry (only keep what we need for visualization) result_entry = { "dataset": dataset_name, "model": model_name, "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, # Store minimal sample results for inspection "sample_results_summary": { "total_samples": len(data.get("sample_results", [])), "first_5_samples": data.get("sample_results", [])[:5] } } 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() return pd.DataFrame(results) 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['ui_type_results'] # For ScreenSpot datasets, we have desktop/web and text/icon if 'screenspot' in dataset_filter.lower(): # Calculate aggregated metrics desktop_text = ui_results.get('desktop_text', {}).get('correct', 0) / max(ui_results.get('desktop_text', {}).get('total', 1), 1) * 100 desktop_icon = ui_results.get('desktop_icon', {}).get('correct', 0) / max(ui_results.get('desktop_icon', {}).get('total', 1), 1) * 100 web_text = ui_results.get('web_text', {}).get('correct', 0) / max(ui_results.get('web_text', {}).get('total', 1), 1) * 100 web_icon = ui_results.get('web_icon', {}).get('correct', 0) / max(ui_results.get('web_icon', {}).get('total', 1), 1) * 100 # Calculate averages desktop_avg = (desktop_text + desktop_icon) / 2 if desktop_text or desktop_icon else 0 web_avg = (web_text + web_icon) / 2 if web_text or web_icon else 0 text_avg = (desktop_text + web_text) / 2 if desktop_text or web_text else 0 icon_avg = (desktop_icon + web_icon) / 2 if desktop_icon or web_icon else 0 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': row['overall_accuracy'] }) 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=400, height=300 ) # 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") # 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 (optional) models = ['All'] + sorted(filtered_df['model'].unique()) selected_model = st.sidebar.selectbox("Select Model", models) if selected_model != 'All': filtered_df = filtered_df[filtered_df['model'] == selected_model] # Main content st.header(f"Results for {selected_dataset}") # 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) if not ui_metrics_df.empty and 'screenspot' in selected_dataset.lower(): st.subheader("Performance by UI Type") # Create charts in a grid col1, col2 = st.columns(2) with col1: # Overall Average chart = create_bar_chart(ui_metrics_df, 'overall', 'Overall Average') if chart: st.altair_chart(chart, use_container_width=True) # Desktop Average chart = create_bar_chart(ui_metrics_df, 'desktop_avg', 'Desktop Average') if chart: st.altair_chart(chart, use_container_width=True) # Text Average chart = create_bar_chart(ui_metrics_df, 'text_avg', 'Text Average (UI-Type)') if chart: st.altair_chart(chart, use_container_width=True) with col2: # Web Average chart = create_bar_chart(ui_metrics_df, 'web_avg', 'Web Average') if chart: st.altair_chart(chart, use_container_width=True) # Icon Average chart = create_bar_chart(ui_metrics_df, 'icon_avg', 'Icon Average (UI-Type)') if chart: st.altair_chart(chart, use_container_width=True) # Detailed breakdown with st.expander("Detailed UI Type Breakdown"): # Create a heatmap-style table detailed_metrics = [] for _, row in ui_metrics_df.iterrows(): detailed_metrics.append({ 'Model': row['model'], 'Desktop Text': f"{row['desktop_text']:.1f}%", 'Desktop Icon': f"{row['desktop_icon']:.1f}%", 'Web Text': f"{row['web_text']:.1f}%", 'Web Icon': f"{row['web_icon']:.1f}%", 'Overall': f"{row['overall']:.1f}%" }) if detailed_metrics: st.dataframe(pd.DataFrame(detailed_metrics), use_container_width=True) else: # For non-ScreenSpot datasets, show a simple bar chart st.subheader("Model Performance") chart_data = filtered_df[['model', 'overall_accuracy']].copy() chart_data.columns = ['Model', 'Accuracy'] chart = alt.Chart(chart_data).mark_bar().encode( x=alt.X('Model:N', sort='-y', axis=alt.Axis(labelAngle=-45)), y=alt.Y('Accuracy:Q', scale=alt.Scale(domain=[0, 100])), tooltip=['Model', 'Accuracy'] ).properties( width=800, height=400 ) st.altair_chart(chart, use_container_width=True) # Model details table with st.expander("Model Details"): display_df = filtered_df[['model', 'overall_accuracy', 'total_samples', 'checkpoint_steps', 'training_loss', 'timestamp']].copy() display_df.columns = ['Model', 'Accuracy (%)', 'Samples', 'Checkpoint Steps', 'Training Loss', 'Timestamp'] display_df['Accuracy (%)'] = display_df['Accuracy (%)'].apply(lambda x: f"{x:.2f}") display_df['Training Loss'] = display_df['Training Loss'].apply(lambda x: f"{x:.4f}" if pd.notna(x) else "N/A") st.dataframe(display_df, use_container_width=True) # Raw data viewer with st.expander("Sample Results"): if selected_model != 'All' and len(filtered_df) == 1: summary = filtered_df.iloc[0]['sample_results_summary'] st.write(f"**Total evaluation samples:** {summary['total_samples']}") st.write("**First 5 sample results:**") for i, sample in enumerate(summary['first_5_samples'], 1): st.write(f"\n**Sample {i}:**") col1, col2 = st.columns([1, 2]) with col1: st.write(f"- **Correct:** {'✅' if sample.get('is_correct') else '❌'}") st.write(f"- **Image:** {sample.get('img_filename', 'N/A')}") with col2: st.write(f"- **Instruction:** {sample.get('instruction', 'N/A')}") if sample.get('predicted_click'): st.write(f"- **Predicted Click:** {sample['predicted_click']}") if sample.get('error_msg'): st.write(f"- **Error:** {sample['error_msg']}") else: st.info("Select a specific model to view sample results") if __name__ == "__main__": main()