Spaces:
Running
Running
Commit History
-
Update UI: Close filters by default and increase quick filter button size Running
Verified signaturedf4bc30 -
Migrate to Gradio app with interactive features
Verified signatureReplace static HTML with Gradio application featuring: - Interactive search and filtering - Model size range slider - Benchmark category filters - Quick filter presets (size and category) - Sortable columns - CSV export - Provider logos and clickable model links - Color-coded scores by benchmark categoryfa808e0
-
Remove authentication requirement - dataset is public
40e90a6 -
Update website to load from parquet dataset
- Replace JSON fetch with parquet dataset loading - Use parquet-wasm and apache-arrow to parse parquet files - Transform flat parquet structure back to nested JSON for UI - Load from https://huggingface.co/api/datasets/OpenEvals/leaderboard-data/parquet - Maintain backward compatibility with existing UI code - Benchmark definitions inline for faster loadinge08bb94 -
Update fetch script to upload parquet dataset to HuggingFace
- Add datasets and pyarrow dependencies to UV script - Flatten nested JSON structure for parquet compatibility - Create columns for all benchmarks (even if null) - Upload directly to OpenEvals/leaderboard-data dataset - Remove local file read/write dependencies - Add aggregate score and coverage metrics - 82 models, 11 benchmarks, 23 columns total16cdd3e -
Add UV script metadata to fetch_api_only.py
- Add PEP 723 header with inline dependencies - Makes script compatible with HuggingFace Scheduled Jobs - Dependencies: requests, huggingface-hub44de261 -
Fetch leaderboard data from private HuggingFace dataset
- Create private dataset OpenEvals/leaderboard-data on HuggingFace - Replace inline LEADERBOARD_DATA (2,718 lines) with fetch call - Load data from https://huggingface.co/datasets/OpenEvals/leaderboard-data/resolve/main/leaderboard.json - Require OAuth authentication to access private dataset - Show loading state while fetching data - Show sign-in prompt if user not authenticated - Reduce index.html from 3,972 to 1,254 lines (68% reduction) - Auto-load data after successful OAuth login155702e
-
Update default page state: deselect GSM8K and sort by TB2
- Remove gsm8k from default selected benchmarks (9 instead of 10) - Set initial sort to Terminal Bench 2.0 (column 10) on page load - Users now see models ranked by TB2 performance by defaultbb4e9a0 -
Match HuggingFace design system and remove About section
- Update fonts to Source Sans Pro and IBM Plex Mono - Update light mode colors to match HF's gray palette - Update dark mode text colors for better contrast - Remove 'About This Leaderboard' section and related CSS - Cleaner, more minimal design matching HF aesthetic0bc1c26 -
Make column sorting more obvious and accessible
- Removed links from benchmark names (were interfering with sorting) - Added small info icon (ℹ️) next to each benchmark name to access benchmark page - Info icons use event.stopPropagation() to prevent triggering sort - Updated tooltips to say 'Click to sort' for clarity - Enhanced hover effect (stronger background, slight lift transform) - Benchmark names are now clearly clickable for sorting - Info icons provide access to benchmark pages without conflict Users can now: - Click column name → Sort by that benchmark - Click ℹ️ icon → Open benchmark page in new tab97cce9e -
Implement benchmark column sorting
- Added full sortTable() function to sort by any column - Clicking a benchmark column sorts models by that benchmark's score - Toggle between descending (↓) and ascending (↑) on repeated clicks - Model name column (index 0) sorts alphabetically - Benchmark columns sort numerically by score - Missing scores (—) always appear at the end - Active sort column highlighted in accent color - Sort arrows update to show current direction (↕ → ↓ → ↑) - Maintains sort state across clicksc67c922 -
Clean up repository - remove unused files
Removed: - 11 PNG screenshot files (1.png - 11.png) - Backup HTML files (benchmarks.html, index-BACK2.html) - Unused text files (llms-full.txt, llms.txt) - SEO files not needed for HF Space (robots.txt, sitemap.xml, schema.jsonld) - Unused CSS file (style.css - CSS is inline in index.html) - Citation file (CITATION.cff) - Extra documentation files (BENCHMARKS_README.md, DATA_UPDATED.md, FIXED.md, IMPLEMENTATION_SUMMARY.md, SUCCESS.md, USAGE_GUIDE.md) - 5 unused Python scripts (curate_model_data.py, fetch_all_benchmarks.py, fetch_from_leaderboards.py, fetch_hle_data.py, populate_real_data.py) Kept: - index.html (main leaderboard page) - README.md (repo documentation) - data/ (leaderboard data and provider logos) - scripts/fetch_api_only.py (fetch benchmark data) - scripts/fetch_provider_logos.py (fetch provider logos) - Git config files (.gitignore, .gitattributes) Result: Clean, minimal repo with only essential files4772857 -
Remove progress bars under benchmark scores
- Removed progress bar (sb/sf divs) from renderScore function - Scores now display as clean numbers with category colors - Simpler, cleaner look focused on the score values - Tooltips still show source and date on hover1ad10a3 -
Fix sticky header alignment to match table position
- Updated sticky header to match tableWrapper's left position and width - Changed from left:0, right:0 to dynamic left and width based on table position - Added table-layout: fixed for consistent column sizing - Header now perfectly aligns with table columns when sticky - Updates position on scroll and resize96eb75a -
Implement JavaScript-based sticky table header
- Created initStickyHeader() function that clones the header - Cloned header is fixed at top of viewport when scrolling - Dynamically matches column widths of original header - Shows/hides based on scroll position - Header appears when original scrolls out of view - Works around overflow-x: auto limitation of pure CSS stickyb40e616 -
Fix sticky table header to stay visible when scrolling
- Increased thead z-index from 10 to 100 for better stacking - Increased model column z-index from 11 to 101 - Added box-shadow to header for better visual separation when floating - Header now stays at top of viewport during page scroll5017b6c -
Use more descriptive benchmark column names
- AIME → AIME'26 (clarifies 2026 version) - HMMT → HMMT Feb'26 (specifies February 2026) - Makes column headers more informative at a glance352fc29 -
Remove aggregate score column from table
- Removed 🏆 Aggregate column header - Removed aggregate score cell from table rows - Models still sorted by aggregate score internally - Updated column indexes for sortTable functions (all shifted by -1) - Updated benchmarkMapping indexes to match new column positions - Cleaner table with more focus on individual benchmark scores5abc218 -
Add SWE-bench Pro and EvasionBench data
- Added ScaleAI/SWE-bench_Pro to fetch script (11 models) - Added FutureMa/EvasionBench to fetch script (5 models) - Total models increased from 67 to 73 - Re-fetched all data with new benchmarks - Updated provider logos (33 organizations, 100% coverage) - Updated LEADERBOARD_DATA in HTML New benchmark coverage: - SWE-Pro: 11 models (Qwen3-Coder-Next leads with 44.3%) - EvasionBench: 5 models (GLM-4.7 leads with 82.91%) Both benchmarks now fully functional in the leaderboard!1f960f8 -
Fix: Remove darker background for top models in name column
- Changed .hl td.c-model background from var(--surface-alt) to var(--surface) - Top models now have same background color as all other models - Uniform appearance across entire table3142a51 -
Remove green background highlight for top models
- Changed .hl class background from green (rgba(22,163,74,.02)) to transparent - Removed green hover effect, now uses standard hover color - Top models (score > 80) no longer have special background color - Cleaner, more uniform table appearance6deb4c6 -
Implement category-based gradient colors for benchmark scores
- Each benchmark category now has a unique color gradient - Math (📐): Purple gradient (light lavender → deep purple) - Knowledge (🧠): Blue gradient (light blue → deep blue) - Coding (💻): Green gradient (light green → dark green) - Agent (🤖): Teal gradient (light cyan → dark teal) - Language (💬): Orange gradient (light orange → dark orange) - Vision (👁️): Pink gradient (light pink → deep rose) - Embedding (🔍): Amber gradient (light yellow → dark amber) - Aggregate: Blue-indigo gradient for overall scores Benefits: - No misleading red/orange colors for low scores on hard benchmarks - Terminal Bench 2.0 score of 43 now shows in teal (neutral) - Each category visually distinct and easy to identify - Removed old grade-based color system (S/A/B/C grades) - All colors based on category, not arbitrary score thresholdsc570da9 -
Remove verified/confidence badges from score cells
- Removed confidence badge rendering from renderScore function - Scores now display cleaner with just the number and progress bar - Source and date info still available in tooltip on hover0c4ff4d -
Fix: Restore provider logos
- Re-ran fetch_provider_logos.py to add providerLogoUrl to all 67 models - All 32 provider organizations have logos (100% success rate) - Updated LEADERBOARD_DATA in HTML with logo URLs - Logos will now display inline with model namesefcfc83 -
Increase table header size
- Font size: 8px → 11px - Padding: 10px 6px → 12px 8px - Letter spacing: .4px → .5px - Better readability and visual hierarchye7ea2bc -
Fix: Use safetensors.total for accurate parameter counts
- Changed from safetensors.parameters.BF16 to safetensors.total - DeepSeek models now show correct sizes (684-685B instead of missing) - Updated slider range to 0-1100B (max model is 1058.6B) - 58/67 models have parameter data (86.6% coverage) - All models now accurately reflect total parameter counts across all precision formatsc04ca45 -
Fix: Restore JavaScript and update LEADERBOARD_DATA with parametersInBillions
- Restore all JavaScript functions that were accidentally removed - Update LEADERBOARD_DATA with parametersInBillions field from API - Fix applyAllFilters to use row.dataset.name instead of wrong selector - Remove debug console.log statementsfe64e89 -
Update LEADERBOARD_DATA with parametersInBillions field
ffc7b47 -
Debug: Add console logging to size filter
18b9793 -
Add model size range slider filter
- Fetch parameter counts from HuggingFace API safetensors field (86.6% coverage, 58/67 models) - Add double-sided range slider (0-800B) in toolbar between search and Sign In - Implement real-time filtering by model size with blue accent styling - Models without size data are always shown - Filter integrates with existing search and benchmark filters (AND logic) - Parse parameter counts from model names as fallback - Update data with parametersInBillions field for all modelscb98dd6 -
Move benchmark filter to separate bar below search
Major Layout Changes: - Move benchmark filter out of toolbar into dedicated div below - Create new .benchmark-filter-bar container with padding and border - Spread categories horizontally using flex with justify-content:space-between - Each category gets flex:1 to distribute evenly across width - Categories use nowrap to stay in single horizontal row - Pills can wrap within their category if needed Visual Improvements: - Categories now spread across full width of container - Better horizontal distribution of space - Cleaner separation between search and filters - Filter bar has its own background/border styling Responsive: - Mobile: Categories stack vertically (flex:1 1 100%) - Action buttons span full width on mobile with top border separator50e5add -
Make categories more compact to fit horizontally without scrollbar
- Reduce category padding (6px 10px -> 4px 8px) - Reduce icon size (14px -> 12px) - Reduce label font size (8.5px -> 8px) - Reduce label min-width (45px -> 38px) - Reduce pill padding (4px 9px -> 3px 7px) - Reduce pill font size (9px -> 8.5px) - Reduce pill border-radius (16px -> 14px) - Allow wrapping but make everything compact enough to minimize it0f25134 -
Force categories to display in single horizontal row
- Change flex-wrap from wrap to nowrap - Add horizontal scrolling if categories overflow (overflow-x:auto) - Categories now stay in one horizontal line instead of stacking - Add padding-bottom for scrollbar clearancef4159b3 -
Make benchmark categories more horizontal and compact
- Reduce gap between categories from 16px to 8px - Make categories flex-shrink:0 to prevent unwanted wrapping - Prevent pills from wrapping within category (nowrap) - Reduce category label font size (9px -> 8.5px) for compactness - Reduce category label min-width (50px -> 45px) - Slightly reduce pill padding (10px -> 9px) - Categories now spread out more horizontally across the toolbar945394e -
Refresh leaderboard data with latest benchmarks
Data Update (2026-03-16 14:50): - Total models: 67 (+3 from previous) - GPQA scores: 27 models (+1) - HLE scores: 29 models (+1) - Provider logos: 32 organizations (+2) New additions: - FireRedBackup and GAIR organizations - 3 new models in various benchmarks - Updated scores from HuggingFace official APIs Benchmarks fetched: - GSM8K: 4 models - MMLU-Pro: 33 models (+1) - GPQA Diamond: 27 models (gated, with auth) - HLE: 29 models (gated, with auth) - SWE-bench Verified: 14 models - AIME 2026: 5 models - HMMT Feb 2026: 5 models - olmOCR: 14 models (+1) - Terminal-Bench 2.0: 16 models (+1)7181682 -
Improve category differentiation in benchmark filter
- Add background boxes around each category (light gray with border) - Increase gap between categories from 12px to 16px - Add vertical separator line before action buttons - Pills now have white background instead of gray for better contrast - Category labels have minimum width for better alignment - Mobile: Add spacing between categories, border-top separator for actions - Much clearer visual separation between Math, Knowledge, Coding, etc.cc40a36 -
Replace dropdown with category-based benchmark filter UI
- Replace dropdown menu with inline category pills (Option B layout) - Benchmarks grouped by category: Math, Knowledge, Coding, Agent, Language, Vision, Embedding - Click pills to toggle on/off with background color change - Visual feedback: blue active state, gray inactive (60% opacity) - Select All/Clear All action buttons on the right - Mobile-friendly: categories wrap, pills flex within category - Default: 10/12 benchmarks selected (Vision & Embedding hidden) - Removed ~60 lines of dropdown code, added ~150 lines for new UI99cd5bf -
Update leaderboard data with GPQA and HLE scores
- Fetched data from gated datasets using HF token - Added GPQA scores for 26 models - Added HLE scores for 28 models - Updated provider logos for 30 organizations - Total: 64 models from official APIs9a5ca50
-
Move provider logo inline with model name
- Remove separate Provider column from table - Add provider logo before model name in same cell - Update column indexes for sorting (Aggregate: 2→1, Benchmarks: 3→2) - Update benchmarkMapping indexes (shift all down by 1) - Add CSS for inline logo (16px, inline-block with margin) - More compact table layout matching original ALL Bench design9763a89 -
Move coverage count under aggregate score like original
- Removed separate Coverage column from table - Coverage now displays as small text under aggregate score - Format: aggregate score (large) with coverage count below (e.g., '8/10') - Matches original ALL Bench leaderboard design - Updated all column indexes after removal - More compact and cleaner layouteb7ca8d -
Replace category buttons with benchmark dropdown
- Removed category filter buttons (Math, Knowledge, Coding, etc.) - Added dropdown menu with checkboxes for all 12 individual benchmarks - Each benchmark has icon and label (📐 GSM8K, 🧠 MMLU-Pro, etc.) - Select All / Deselect All actions in dropdown header - Shows 'X selected' in dropdown button (e.g., '10 selected') - Default: 10 benchmarks selected (excludes Vision/OCR and Embedding) - Checkboxes update in real-time - Filters models and columns based on selected benchmarks - Styled dropdown with smooth animations and dark mode support - Click outside to close dropdownda2d10b -
Fix aggregate scoring - use original scores instead of recalculating
Issue: Models with single high score were ranking higher than models with many good scores Solution: Keep original aggregate scores for ranking, only filter visibility Changes: - Aggregate score now shows the model's overall performance (original calculation) - Models are ranked by their full benchmark average, not just visible ones - This prevents single-benchmark models from artificially ranking higher - Coverage count still shows X/Y for visible benchmarks only - Only models with scores in visible categories are shown Example: - Model A: 95 score in 1 benchmark (GSM8K only) - Model B: 85 average across 8 benchmarks - Model B now correctly ranks higher due to broader performance854e6fd -
Add dynamic aggregate score calculation and smart model filtering
- Aggregate scores now recalculate based on visible benchmarks only - Models without scores in visible benchmarks are hidden - Coverage count updates to show X/Y where Y is total visible benchmarks - Stats (model count, total scores) update dynamically per category - Page loads with Math, Knowledge, Coding, Agent, Language visible by default - OCR and Embedding hidden by default but can be shown - Clicking category buttons filters both models and recalculates scores - Rankings update based on filtered benchmarks onlyb34b439 -
Add category filters and remove Open Source button
- Remove Open Source filter (all models are open source) - Add category filter buttons: Math, Knowledge, Coding, Vision, Embedding, Agent, Language - Rename 'retrieval' category to 'embedding' - Hide/show benchmark columns based on selected category - Default: hide OCR (vision) and Embedding columns on page load - Update CSS class from bm-retrieval to bm-embedding - Sync both index.html and benchmarks.htmlbf1bdd9 -
Fix OAuth scopes in README metadata
f52d675 -
Add official benchmarks leaderboard with OAuth support
- Unified leaderboard for 12 official HuggingFace benchmarks - OAuth authentication for gated datasets (GPQA, HLE) - Provider logos fetched from HuggingFace API - Clickable model names linking to HF model pages - Real-time data from official leaderboard APIs - Beautiful gradient UI with dark mode - 53 open-source models with 80+ scores - Interactive search and filtering - Responsive design58e6611 -
Update index.html
Verified signaturef8278c5 -
Update index-BACK2.html
Verified signatureadbdb07 -
Update index.html
Verified signatureade638e -
Update index.html
Verified signature791df73
