Spaces:
Running
Running
| import pandas as pd | |
| import numpy as np | |
| import gradio as gr | |
| from urllib.parse import quote | |
| def style_dataframe(df): | |
| if len(df) == 0: | |
| return df | |
| highlight_cols = ["Average", "IFEval", "BBHI", "MATH", "GPQA", "MUJB", "MMLU-PRO"] | |
| styled = df.style | |
| def highlight_green(val): | |
| try: | |
| val_float = float(str(val).replace('%', '').replace(' kg', '')) | |
| return f'background: linear-gradient(90deg, rgba(46, 125, 50, 0.5) {val_float}%, rgba(46, 125, 50, 0.1) {val_float}%); color: white;' | |
| except: | |
| return 'background-color: #121212; color: white;' | |
| for col in highlight_cols: | |
| styled = styled.applymap(highlight_green, subset=[col]) | |
| styled = styled.set_properties( | |
| subset=["Model"], | |
| **{'color': '#4da6ff'} | |
| ) | |
| return styled | |
| def create_leaderboard_data(): | |
| models = [ | |
| {"model_name": "meta-llama/llama-3-70b-instruct", "type": "open"}, | |
| {"model_name": "mistralai/Mistral-7B-Instruct-v0.3", "type": "open"}, | |
| {"model_name": "google/gemma-7b-it", "type": "open"}, | |
| {"model_name": "Qwen/Qwen2-7B-Instruct", "type": "open"}, | |
| {"model_name": "anthropic/claude-3-opus", "type": "closed"}, | |
| {"model_name": "OpenAI/gpt-4o", "type": "closed"}, | |
| {"model_name": "01-ai/Yi-1.5-34B-Chat", "type": "open"}, | |
| {"model_name": "google/gemma-2b", "type": "open"}, | |
| {"model_name": "microsoft/phi-3-mini-4k-instruct", "type": "open"}, | |
| {"model_name": "microsoft/phi-3-mini-128k-instruct", "type": "open"}, | |
| {"model_name": "stabilityai/stable-beluga-7b", "type": "open"}, | |
| {"model_name": "togethercomputer/RedPajama-INCITE-7B-Instruct", "type": "open"}, | |
| {"model_name": "databricks/dbrx-instruct", "type": "closed"}, | |
| {"model_name": "mosaicml/mpt-7b-instruct", "type": "open"}, | |
| {"model_name": "01-ai/Yi-1.5-9B-Chat", "type": "open"} | |
| ] | |
| np.random.seed(42) | |
| rows = [] | |
| for i, model in enumerate(models, 1): | |
| model_name = model["model_name"] | |
| model_type = model["type"] | |
| emoji = "🟢" if model_type.lower() == "open" else "🔴" | |
| type_with_emoji = f"{emoji} {model_type.upper()}" | |
| if "/" in model_name: | |
| org, name = model_name.split("/", 1) | |
| model_link = f"[{model_name}](https://huggingface.co/{quote(model_name)})" | |
| else: | |
| model_link = f"[{model_name}](https://huggingface.co/models?search={quote(model_name)})" | |
| average = round(np.random.uniform(40, 90), 2) | |
| ifeval = round(np.random.uniform(30, 90), 2) | |
| bbhi = round(np.random.uniform(40, 85), 2) | |
| math = round(np.random.uniform(20, 80), 2) | |
| gpqa = round(np.random.uniform(10, 70), 2) | |
| mujb = round(np.random.uniform(10, 70), 2) | |
| mmlu = round(np.random.uniform(40, 85), 2) | |
| co2_cost = round(np.random.uniform(1, 100), 2) | |
| rows.append([ | |
| i, | |
| type_with_emoji, | |
| model_link, | |
| f"{average}", | |
| f"{ifeval}", | |
| f"{bbhi}", | |
| f"{math}", | |
| f"{gpqa}", | |
| f"{mujb}", | |
| f"{mmlu}", | |
| f"{co2_cost} kg" | |
| ]) | |
| rows.sort(key=lambda x: float(x[3]), reverse=True) | |
| for i, row in enumerate(rows, 1): | |
| row[0] = i | |
| df = pd.DataFrame(rows, columns=["Rank", "Type", "Model", "Average", "IFEval", "BBHI", "MATH", "GPQA", "MUJB", "MMLU-PRO", "CO_Cost"]) | |
| return style_dataframe(df) | |
| def get_filter_data(): | |
| return { | |
| "For Edge Devices": 5, | |
| "For Consumers": 4, | |
| "Mid-range": 4, | |
| "For the GPU-rich": 3, | |
| "Only Official Providers": 8 | |
| } | |
| css = """ | |
| .html-container { | |
| text-align: center; | |
| display: flex; | |
| justify-content: center; | |
| width: 100%; | |
| } | |
| .dataframe-container { | |
| margin-top: 0.5rem; | |
| margin-bottom: 0.5rem; | |
| } | |
| .leaderboard-title { | |
| font-size: 1.5rem; | |
| font-weight: bold; | |
| margin-bottom: 0.25rem; | |
| color: #f0f0f0; | |
| } | |
| .leaderboard-subtitle { | |
| font-size: 0.9rem; | |
| margin-bottom: 1rem; | |
| color: #a0a0a0; | |
| } | |
| .filters-container { | |
| margin-bottom: 0.5rem; | |
| } | |
| .hf-logo { | |
| height: 1.5rem; | |
| margin-right: 0.5rem; | |
| } | |
| .header-container { | |
| display: flex; | |
| align-items: center; | |
| justify-content: center; | |
| margin-bottom: 10px; | |
| } | |
| .header-container svg { | |
| margin-right: 0.5rem; | |
| flex-shrink: 0; | |
| } | |
| """ | |
| filter_data = get_filter_data() | |
| filter_choices = [f"{key} · {value}" for key, value in filter_data.items()] | |
| # HF logo SVG content to embed directly | |
| hf_logo_svg = """ | |
| <svg xmlns="http://www.w3.org/2000/svg" width="60" height="60" version="1.1" viewBox="0 0 236 220"> | |
| <g fill="none" stroke="none" stroke-width="1"><path d="M235.188 161.726a21.794 21.794 0 0 0-2.962-6.55 25.09 25.09 0 0 0 .654-3.21c.953-6.877-1.43-13.172-5.924-18.043-2.434-2.66-5.042-4.417-7.805-5.495a104.244 104.244 0 0 0 2.74-23.667c.006-3.638-.199-7.203-.545-10.715a125.62 125.62 0 0 0-.652-5.248 104.85 104.85 0 0 0-3.512-15.2 105.286 105.286 0 0 0-3.513-9.635 105.803 105.803 0 0 0-6.951-13.524 93.949 93.949 0 0 0-5.648-8.355 79.37 79.37 0 0 0-3.113-3.965 104 104 0 0 0-10.394-10.916 79.529 79.529 0 0 0-3.814-3.292 94.49 94.49 0 0 0-3.94-3.112 110.693 110.693 0 0 0-8.358-5.646C155.582 5.522 136.986 0 117.111 0 59.239 0 12.334 46.902 12.334 104.766a104.258 104.258 0 0 0 2.786 23.96c-2.486 1.08-4.87 2.76-7.077 5.194-4.493 4.868-6.877 11.142-5.923 18.017.144 1.093.362 2.174.652 3.238a21.868 21.868 0 0 0-2.962 6.55c-1.38 5.249-.926 9.98.83 14.127-1.908 5.42-1.43 11.193 1.053 16.21 1.807 3.665 4.392 6.5 7.58 9.034 3.788 3.01 8.531 5.57 14.254 8.026 6.826 2.91 15.157 5.646 18.947 6.65 9.788 2.534 19.173 4.14 28.686 4.217 13.55.127 25.222-3.06 33.578-11.215 4.105.503 8.236.755 12.372.754 4.37-.01 8.734-.286 13.07-.824 8.336 8.206 20.051 11.422 33.653 11.292 9.51-.075 18.896-1.68 28.659-4.22 3.814-1.002 12.143-3.738 18.973-6.647 5.723-2.47 10.465-5.029 14.28-8.03 3.163-2.54 5.746-5.37 7.554-9.04 2.51-5.019 2.962-10.79 1.08-16.21 1.727-4.146 2.178-8.889.798-14.133Zm-9.71 13.776c1.922 3.648 2.045 7.77.348 11.61-2.573 5.82-8.964 10.403-21.377 15.325-7.718 3.061-14.787 5.017-14.85 5.036-10.22 2.647-19.453 3.991-27.446 3.991-13.234 0-23.082-3.65-29.33-10.863a98.39 98.39 0 0 1-31.93.18c-6.251 7.093-16.045 10.683-29.17 10.683-7.993 0-17.225-1.344-27.435-3.99-.063-.02-7.127-1.976-14.85-5.037-12.41-4.922-18.803-9.502-21.375-15.325-1.697-3.84-1.574-7.962.35-11.61.177-.339.367-.67.57-.989A14.734 14.734 0 0 1 7 162.613c.766-2.902 2.344-5.316 4.487-7.05a14.685 14.685 0 0 1-1.99-5.548c-.618-4.273.8-8.54 3.992-12.02 2.485-2.707 5.998-4.197 9.89-4.197h.102a97.174 97.174 0 0 1-4.41-29.036c0-53.702 43.54-97.24 97.252-97.24 53.714 0 97.253 43.535 97.253 97.24a97.176 97.176 0 0 1-4.44 29.11c.468-.046.929-.07 1.38-.07 3.89 0 7.404 1.489 9.888 4.196 3.19 3.476 4.61 7.747 3.993 12.02a14.685 14.685 0 0 1-1.99 5.549c2.142 1.733 3.72 4.147 4.485 7.05a14.749 14.749 0 0 1-1.99 11.9c.206.315.398.646.574.985Z" fill="#FFF" fill-rule="nonzero"/><path d="M225.53 174.038a14.752 14.752 0 0 0 1.99-11.903c-.765-2.903-2.345-5.318-4.489-7.052a14.7 14.7 0 0 0 1.992-5.55c.617-4.274-.801-8.543-3.996-12.023-2.486-2.708-6.002-4.198-9.894-4.198-.453 0-.912.023-1.382.07a97.193 97.193 0 0 0 4.435-29.116C214.186 50.55 170.616 7 116.873 7s-97.317 43.546-97.317 97.266a97.164 97.164 0 0 0 4.413 29.044h-.103c-3.892 0-7.408 1.49-9.894 4.198-3.192 3.477-4.613 7.748-3.996 12.023.282 1.966.96 3.853 1.992 5.55-2.144 1.734-3.724 4.15-4.487 7.053a14.734 14.734 0 0 0 1.992 11.904c-.206.318-.395.65-.572.988-1.923 3.65-2.048 7.774-.35 11.614 2.574 5.82 8.971 10.407 21.392 15.33 7.723 3.061 14.797 5.018 14.86 5.036C55.017 209.656 64.256 211 72.254 211c13.134 0 22.933-3.59 29.192-10.686a98.495 98.495 0 0 0 31.951-.181c6.25 7.214 16.105 10.867 29.35 10.867 7.998 0 17.238-1.345 27.454-3.994.062-.018 7.132-1.975 14.859-5.037 12.42-4.922 18.818-9.508 21.391-15.33 1.699-3.84 1.575-7.963-.349-11.613a11.9 11.9 0 0 0-.573-.988Zm-128.562 15.17a39.223 39.223 0 0 1-1.73 2.756c-1.618 2.367-3.746 4.175-6.215 5.52-4.722 2.573-10.697 3.473-16.77 3.473-9.59 0-19.42-2.243-24.931-3.673-.27-.07-33.778-9.53-29.536-17.58.714-1.355 1.89-1.896 3.368-1.896 5.973 0 16.849 8.89 21.524 8.89 1.043 0 1.78-.445 2.08-1.528 1.992-7.142-30.276-10.144-27.558-20.484.48-1.83 1.78-2.572 3.61-2.572 7.895 0 25.62 13.88 29.326 13.88.284 0 .488-.084.599-.259l.048-.08c1.739-2.87.742-4.958-11.18-12.255l-1.143-.697c-13.116-7.935-22.323-12.71-17.086-18.407.602-.658 1.455-.95 2.493-.95 1.23 0 2.72.413 4.36 1.102 6.927 2.92 16.528 10.884 20.539 14.357 1.182 1.03 1.879 1.657 1.879 1.657s5.078 5.28 8.15 5.28c.706 0 1.305-.278 1.71-.967 2.18-3.67-20.222-20.636-21.486-27.636-.866-4.745.59-7.147 3.286-7.147 1.279 0 2.844.545 4.569 1.64 5.355 3.397 15.693 21.153 19.478 28.06 1.268 2.316 3.434 3.294 5.386 3.294 3.872 0 6.9-3.847.354-8.737-9.837-7.357-6.385-19.382-1.69-20.122.2-.032.403-.048.604-.049 4.27 0 6.154 7.355 6.154 7.355s5.52 13.855 15.003 23.325c8.605 8.596 9.806 15.651 4.793 24.45Zm30.664 1.622-.492.059-.84.096c-.44.045-.881.089-1.325.13l-.433.04-.394.033-.56.045c-.206.015-.413.03-.62.043l-.616.042-.136.008-.486.029-.206.01-.576.029-.67.027-.607.02-.404.013h-.205c-.127 0-.252.007-.38.008h-.2c-.127 0-.252 0-.38.007l-.514.005h-.72c-.566 0-1.13-.005-1.692-.016l-.456-.01c-.13 0-.26-.007-.388-.01l-.482-.015-.6-.023-.54-.023-.141-.006-.515-.028c-.143-.008-.286-.015-.427-.024l-.333-.02a91.779 91.779 0 0 1-1.254-.088l-.438-.035-.552-.047-.646-.06a57.45 57.45 0 0 1-1.082-.11h-.017c5.263-11.734 2.6-22.692-8.033-33.313-6.973-6.96-11.614-17.242-12.575-19.5-1.95-6.68-7.106-14.107-15.67-14.107a14.01 14.01 0 0 0-2.161.171c-3.753.59-7.032 2.75-9.372 5.997-2.528-3.144-4.986-5.642-7.21-7.055-3.35-2.122-6.692-3.2-9.951-3.2-4.065 0-7.7 1.67-10.233 4.697l-.064.077-.142-.6-.006-.025a85.253 85.253 0 0 1-1.214-6.23c0-.014 0-.028-.007-.042l-.071-.473c-.04-.255-.076-.512-.113-.767-.03-.213-.06-.426-.088-.638-.027-.212-.057-.425-.082-.638a64.55 64.55 0 0 1-.15-1.247l-.006-.052a89.658 89.658 0 0 1-.267-2.807l-.04-.542a27.47 27.47 0 0 1-.031-.45l-.017-.101-.037-.601a47.596 47.596 0 0 1-.037-.653 58.953 58.953 0 0 1-.032-.652l-.023-.57-.008-.173-.017-.527-.013-.446c0-.18-.008-.358-.012-.537l-.01-.563c-.001-.187 0-.374-.003-.561-.006-.18 0-.37 0-.56 0-48.17 39.07-87.22 87.268-87.22 48.197 0 87.268 39.05 87.268 87.225v1.127c0 .188-.005.377-.01.563 0 .157-.005.31-.01.466-.001.14-.007.279-.013.413 0 .176-.01.35-.016.527v.015c-.016.383-.031.765-.048 1.147l-.007.126-.03.57-.027.455a90.924 90.924 0 0 1-.251 3.212v.014l-.06.602-.155 1.398-.058.469-.073.55c-.025.2-.053.4-.082.603-.03.225-.064.448-.097.67l-.08.529-.095.598c-.032.2-.064.397-.102.596-.04.198-.07.396-.104.595a60.2 60.2 0 0 1-.217 1.186 104.22 104.22 0 0 1-.612 2.935c-2.457-2.386-5.71-3.685-9.297-3.685-3.255 0-6.602 1.076-9.951 3.2-2.225 1.41-4.681 3.91-7.211 7.054-2.343-3.25-5.623-5.408-9.37-5.998a14.168 14.168 0 0 0-2.163-.17c-8.566 0-13.72 7.428-15.67 14.109-.967 2.257-5.607 12.537-12.59 19.51-10.626 10.586-13.31 21.495-8.127 33.18Zm90.093-23.488-.035.102a5.71 5.71 0 0 1-.074.196 6.04 6.04 0 0 1-.51.95c-.192.286-.402.56-.632.817-.053.06-.103.12-.163.178a4.67 4.67 0 0 1-.25.26c-1.55 1.537-3.91 2.883-6.58 4.12-.301.136-.61.273-.92.41l-.31.137c-.203.09-.41.182-.625.27-.21.09-.424.179-.635.266l-.643.266c-1.503.618-3.042 1.206-4.545 1.789l-.64.248-.638.249c-.42.164-.838.329-1.25.494l-.61.246-.6.246-.296.123c-.195.083-.389.164-.58.247-4.418 1.894-7.597 3.822-6.93 6.215.02.068.04.132.064.194.06.177.14.345.242.502.06.094.13.183.207.264.785.814 2.214.686 4.016.039.147-.053.297-.11.448-.17.103-.039.206-.08.31-.121l.155-.066a33.9 33.9 0 0 0 1.295-.582c.113-.052.226-.103.339-.163 2.203-1.076 4.7-2.527 7.178-3.847.38-.206.76-.403 1.14-.6.62-.32 1.246-.63 1.877-.925 2.347-1.103 4.552-1.886 6.326-1.886.832 0 1.568.17 2.18.562l.1.068c.31.22.58.487.8.795.05.073.104.148.149.225.049.08.095.16.138.244.877 1.662.143 3.385-1.573 5.073-1.647 1.62-4.21 3.21-7.125 4.682-.215.11-.433.219-.652.325-8.672 4.256-20.026 7.46-20.184 7.5-3.027.785-7.354 1.815-12.228 2.585l-.722.112-.118.017a94.839 94.839 0 0 1-3.325.439l-.104.01c-2.02.23-4.091.397-6.167.467h-.031c-.747.028-1.492.04-2.238.04h-.86a54.33 54.33 0 0 1-3.407-.154c-.027 0-.054 0-.082-.005a43.83 43.83 0 0 1-3.298-.39c-.27-.043-.557-.09-.835-.14a19.884 19.884 0 0 1-.377-.07l-.03-.006c-.4-.076-.797-.16-1.193-.25-.23-.053-.46-.104-.687-.165l-.135-.036a14.137 14.137 0 0 1-.334-.088l-.061-.016-.356-.102c-.131-.037-.258-.075-.387-.113l-.046-.014-.336-.102-.378-.12c-.103-.035-.206-.071-.31-.105l-.226-.08a22.09 22.09 0 0 1-.652-.24l-.207-.078-.17-.068a27.61 27.61 0 0 1-.971-.412l-.214-.102-.036-.016c-.076-.036-.151-.072-.227-.104a19.743 19.743 0 0 1-.442-.22l-.044-.022-.212-.111a17.218 17.218 0 0 1-1.108-.63l-.198-.12c-.1-.062-.192-.125-.29-.19l-.258-.17a9.118 9.118 0 0 1-.278-.19l-.167-.118a15.402 15.402 0 0 1-.515-.386c-.09-.07-.18-.14-.268-.206-.196-.157-.39-.32-.58-.487l-.007-.006c-.185-.16-.366-.326-.542-.495l-.01-.012a11.813 11.813 0 0 1-.271-.266c-.09-.098-.18-.188-.265-.28a12.109 12.109 0 0 1-.26-.28 10.55 10.55 0 0 1-.246-.274l-.028-.03a17.27 17.27 0 0 1-1.165-1.514l-.143-.21c-.094-.14-.188-.28-.28-.42l-.272-.419a34.93 34.93 0 0 1-.66-1.059c-.074-.127-.15-.252-.22-.376a8.379 8.379 0 0 1-.117-.205c-.038-.073-.083-.145-.123-.218l-.066-.115c-.02-.038-.028-.05-.04-.075a12.086 12.086 0 0 1-.228-.417 8.435 8.435 0 0 0-.102-.182l-.104-.198-.101-.197-.198-.4a33.357 33.357 0 0 1-.291-.617c-.026-.056-.053-.112-.076-.169a13.103 13.103 0 0 1-.254-.583c-.055-.13-.104-.26-.157-.388-.15-.38-.288-.763-.416-1.15-.038-.116-.076-.23-.11-.345a13.436 13.436 0 0 1-.224-.79 13.83 13.83 0 0 1-.216-.94 13.172 13.172 0 0 1-.181-1.113 14.8 14.8 0 0 1-.1-1.105c0-.06-.006-.123-.007-.184a9.707 9.707 0 0 1-.01-.366c-.065-4.917 2.425-9.646 7.749-14.964 9.48-9.47 15.002-23.325 15.002-23.325s.148-.58.458-1.417c.103-.277.212-.554.326-.826.13-.309.27-.614.42-.913l.044-.088c.192-.382.4-.753.627-1.116.052-.084.103-.168.163-.25.168-.252.346-.495.534-.73.145-.18.3-.356.462-.522.55-.566 1.179-1.029 1.889-1.276l.088-.029a3.025 3.025 0 0 1 .394-.104l.033-.007c.15-.03.302-.05.455-.06h.014c.08 0 .158-.01.24-.01.204.001.407.017.608.05.855.133 1.67.642 2.37 1.427l.093.11c.228.267.436.553.62.852.138.22.268.454.39.7.05.103.098.197.145.3.023.05.046.101.068.154a11.557 11.557 0 0 1 .69 2.216c.14.703.22 1.414.242 2.13.01.313.01.63 0 .953-.005.122-.011.245-.02.37a13.766 13.766 0 0 1-.885 3.995l-.155.384a14.23 14.23 0 0 1-.86 1.718 15.28 15.28 0 0 1-.6.94l-.13.186a15.584 15.584 0 0 1-1.695 1.99 17.893 17.893 0 0 1-1.956 1.692 15.79 15.79 0 0 0-1.948 1.7c-1.73 1.815-2.133 3.416-1.744 4.63.046.137.1.272.163.402.14.282.33.54.55.766l.063.061c.064.061.13.12.198.176l.07.058c.168.13.345.246.53.346.053.028.103.057.164.085.254.126.52.228.794.305l.075.02.103.027.09.023.095.02.096.02.09.015c.07.013.138.023.206.032l.067.01.118.013.074.007.122.01h.07l.127.005h.4l.113-.006.132-.008.16-.015.15-.018c.03 0 .063-.01.1-.015a5.824 5.824 0 0 0 1.825-.585 6.043 6.043 0 0 0 1.335-.92l.162-.154c.026-.025.052-.049.076-.075.16-.16.31-.327.448-.503a6.87 6.87 0 0 0 .655-.99 264.026 264.026 0 0 1 7.083-12.099l.34-.542.341-.543c.171-.274.343-.546.515-.814l.173-.27.347-.54c.464-.72.931-1.43 1.4-2.135l.352-.525a123.949 123.949 0 0 1 2.103-3.023l.348-.477a63.365 63.365 0 0 1 2.364-3.063l.325-.389c.053-.065.103-.129.161-.192l.32-.368c.052-.06.103-.12.158-.18l.313-.349.154-.168c.15-.162.302-.322.456-.48.104-.103.2-.207.3-.302a12.036 12.036 0 0 1 1.921-1.597l.16-.102.163-.104a6.65 6.65 0 0 1 .313-.185c2.724-1.543 4.978-1.657 6.275-.36.786.786 1.22 2.087 1.197 3.898 0 .08 0 .16-.007.243v.09c0 .081-.007.165-.012.25 0 .102-.013.205-.023.308-.01.103-.015.182-.026.273 0 .031-.005.056-.009.084a6.98 6.98 0 0 1-.029.241c0 .025 0 .05-.01.075a5.115 5.115 0 0 1-.047.326c-.013.1-.03.2-.049.3-.008.058-.02.115-.028.172-.03.164-.07.326-.117.485a6.95 6.95 0 0 1-.324.871 12.3 12.3 0 0 1-.174.367c-.125.25-.265.506-.42.77-.077.13-.156.265-.24.4l-.13.204c-.13.207-.268.415-.41.625a30.251 30.251 0 0 1-1.136 1.528l-.18.228a61.886 61.886 0 0 1-1.56 1.864l-.423.478-.216.241c-.289.33-.585.652-.882.976l-.228.246c-.15.163-.309.33-.46.494-.154.167-.31.332-.467.498l-.474.5-.48.5-.486.503-.978 1.008c-4.67 4.787-9.59 9.558-11.255 12.532-.11.194-.21.393-.3.597-.237.54-.336 1.004-.27 1.378.023.123.067.241.131.348.094.163.208.312.34.446.06.06.123.116.193.167.342.244.756.37 1.176.356h.133l.134-.009.136-.018.112-.016a.307.307 0 0 0 .045-.008l.102-.02.028-.007.115-.024.04-.01.12-.03.145-.043a6.165 6.165 0 0 0 .894-.35c.05-.025.103-.048.152-.074l.153-.076a14.039 14.039 0 0 0 1.078-.623l.154-.102a2.39 2.39 0 0 0 .153-.104l.154-.102.08-.056.22-.154c.207-.143.4-.289.595-.437l.018-.013.308-.238c.423-.333.825-.67 1.188-.986l.24-.213.024-.02.126-.112c.297-.266.562-.515.783-.72l.09-.088.217-.206.13-.13.045-.045.014-.012.22-.223.01-.007.04-.037.05-.044.017-.015.044-.04.234-.206.13-.118a5.41 5.41 0 0 0 .207-.187l.158-.138c.028-.023.055-.05.084-.074l.165-.144.243-.212.128-.112c.5-.436 1.11-.962 1.81-1.555l.286-.24.475-.395.483-.406c.634-.524 1.316-1.083 2.036-1.663l.474-.378 1.239-.977c.168-.132.34-.264.515-.396a132.43 132.43 0 0 1 4.333-3.208l.443-.308c.309-.214.618-.427.927-.635l.28-.19a72.059 72.059 0 0 1 1.675-1.094l.28-.177.278-.173c.278-.172.556-.342.83-.507l.275-.163.552-.32c.18-.107.36-.209.54-.31l.108-.06.428-.235c.178-.097.356-.19.532-.28l.265-.136.256-.13c.585-.293 1.18-.565 1.786-.814l.246-.104a8.13 8.13 0 0 1 .484-.184c.16-.053.318-.109.473-.164.142-.047.28-.09.412-.133l.045-.013a4.62 4.62 0 0 1 .214-.06l.02-.007c.146-.042.29-.08.438-.114h.01a9.31 9.31 0 0 1 .63-.13 7.52 7.52 0 0 1 .808-.097 5.27 5.27 0 0 1 .38-.012h.098a4.179 4.179 0 0 1 .545.042h.023c.055.005.11.015.165.026.055.012.11.02.164.032h.018c.053.012.103.027.158.042.44.118.844.342 1.18.65l.029.03.059.058.057.06a6.346 6.346 0 0 1 1.132 1.647l.045.103a3.788 3.788 0 0 1 .112 2.76 5.722 5.722 0 0 1-.36.832l-.094.17a9.51 9.51 0 0 1-1.265 1.672l-.104.11a14.43 14.43 0 0 1-.942.913l-.122.11c-.253.225-.513.444-.775.658a28.872 28.872 0 0 1-.706.563 49.385 49.385 0 0 1-2.062 1.492 81.87 81.87 0 0 1-1.277.855c-.82.537-1.646 1.063-2.48 1.58-.24.147-.482.298-.73.448-2.264 1.393-4.775 2.882-7.453 4.501l-.694.421c-.76.463-1.476.905-2.15 1.325l-.34.214-.641.412c-.256.161-.504.32-.744.477a33.2 33.2 0 0 0-.533.343l-.343.226a47.78 47.78 0 0 0-.5.33l-.16.101c-.17.113-.337.225-.498.335l-.266.18-.31.215-.286.2c-.45.317-.897.641-1.34.97l-.153.117a31.2 31.2 0 0 0-.505.39l-.21.17a23.3 23.3 0 0 0-.973.823l-.14.131c-.082.075-.163.15-.24.224-.053.05-.102.102-.156.153l-.075.073c-.166.165-.326.334-.481.508l-.076.087a7.57 7.57 0 0 0-.447.567l-.058.083c-.15.218-.282.45-.395.69l-.04.084-.025.06-.02.05-.026.07c-.008.024-.024.067-.034.103a2.517 2.517 0 0 0-.104.423l-.012.077-.01.07-.005.068v.361c0 .03 0 .061.007.092a.899.899 0 0 0 .006.055c0 .03.008.06.01.09l.023.127v.006c.008.04.016.08.027.12.01.041.02.086.033.13a3.7 3.7 0 0 0 .135.383c0 .01.008.02.01.03.015.03.028.062.043.093a4.4 4.4 0 0 0 .189.385l.072.13.076.13.08.134a.47.47 0 0 0 .046.06l.027.03.029.024.03.024a.644.644 0 0 0 .066.04.872.872 0 0 0 .157.056c.664.151 2.03-.399 3.844-1.357l.328-.173.55-.299.27-.149c.193-.104.39-.219.589-.333l.365-.206c2.399-1.38 5.265-3.153 8.226-4.884.277-.163.555-.324.834-.485l.56-.357c.28-.162.558-.32.838-.477a91.851 91.851 0 0 1 3.617-1.95l.547-.277c.364-.18.724-.355 1.08-.525.69-.327 1.386-.637 2.09-.93l.386-.157.047-.016c2.043-.81 3.895-1.31 5.395-1.31a5.42 5.42 0 0 1 .968.08h.01c.103.018.196.04.29.063h.018c.138.033.274.076.405.13.105.038.206.083.304.136a2.89 2.89 0 0 1 1.137 1.038c.1.154.189.316.26.487.043.098.082.2.119.308.036.11.069.217.102.33.244.873.22 1.796-.07 2.655Z" fill="#FF9D00" fill-rule="nonzero"/><path d="M204.17 104.828v-.565C204.17 56.07 165.098 17 116.882 17c-48.214 0-87.3 39.071-87.3 87.263l-.002.19v.375c.006.187.008.375.007.563l.007.414.002.15.006.205c.002.11.005.22.005.33l.012.448.018.527.006.172.023.545v.025l.032.637v.016c.012.218.023.436.037.652.012.201.024.402.038.602l.004.036c.01.171.021.344.034.516l.004.045c.011.166.024.332.038.497l.002.032.023.288c.07.83.15 1.66.242 2.49l.006.05.07.608.08.64a78.863 78.863 0 0 0 .281 2.045l.005.03a84.93 84.93 0 0 0 1.288 6.717l.006.027.037.155c.034.148.07.295.106.444l.06-.078c2.53-3.028 6.17-4.698 10.231-4.698 3.26 0 6.605 1.077 9.956 3.201 2.224 1.412 4.682 3.913 7.211 7.058 2.341-3.25 5.622-5.41 9.375-6a14.05 14.05 0 0 1 2.163-.17c8.568 0 13.727 7.43 15.676 14.114.962 2.26 5.604 12.543 12.603 19.5 10.64 10.625 13.3 21.59 8.037 33.33h.018c.36.04.72.08 1.081.11l.646.07.076.006c.158.015.317.028.476.04l.438.037c.417.032.835.06 1.256.088l.33.021.265.015.165.009.514.028.142.005.542.023.6.024.485.014.08.004c.102.003.206.008.309.008l.11.002c.679.016 1.357.024 2.037.024h.72l.517-.007c.127-.006.252-.006.379-.006h.196l.174-.004a6.09 6.09 0 0 1 .205-.004h.207l.404-.012.608-.02.67-.03c.192-.007.383-.016.575-.026l.207-.012.305-.017.18-.01.134-.008.618-.042c.206-.012.413-.03.62-.045l.56-.043.394-.034.433-.04c.443-.04.885-.086 1.326-.13.28-.03.56-.064.84-.098l.49-.059c-5.18-11.69-2.5-22.606 8.097-33.19 6.983-6.974 11.625-17.26 12.593-19.518 1.95-6.684 7.106-14.115 15.675-14.115.724.001 1.448.058 2.163.172 3.75.59 7.032 2.75 9.376 6 2.53-3.146 4.987-5.647 7.21-7.06 3.351-2.123 6.7-3.2 9.955-3.2 3.59 0 6.843 1.3 9.3 3.687.135-.6.265-1.204.39-1.807.076-.377.15-.754.224-1.133.09-.465.175-.933.254-1.4.02-.125.042-.252.066-.38.038-.198.072-.395.103-.595l.013-.077.08-.525.082-.528v-.006c.033-.22.067-.443.097-.666l.07-.505.013-.097.07-.55.06-.47c.036-.307.07-.616.104-.926.015-.123.027-.247.04-.372l.011-.1c.023-.206.043-.41.06-.617l.046-.479c.096-1.06.173-2.122.233-3.185v-.007l.03-.564.006-.126.023-.536c.01-.2.017-.405.024-.61v-.016l.006-.15a7.99 7.99 0 0 0 .01-.378l.003-.05c.004-.118.01-.24.01-.36l.002-.104c.004-.12.007-.24.007-.362l.003-.102c.003-.155.006-.308.006-.462v-.562ZM95.23 192.004c6.914-10.136 6.424-17.742-3.062-27.22-9.486-9.476-15.007-23.335-15.007-23.335s-2.061-8.052-6.76-7.311c-4.7.74-8.148 12.769 1.691 20.132 9.84 7.364-1.957 12.357-5.744 5.447-3.789-6.911-14.129-24.675-19.49-28.073-5.362-3.398-9.132-1.494-7.868 5.51.627 3.476 6.465 9.408 11.92 14.952 5.534 5.625 10.672 10.847 9.575 12.695-2.181 3.67-9.86-4.312-9.86-4.312s-24.048-21.875-29.284-16.174c-4.825 5.254 2.616 9.72 14.085 16.604.975.586 1.98 1.19 3.008 1.81 13.12 7.942 14.141 10.037 12.28 13.04-.688 1.11-5.08-1.525-10.483-4.767-9.198-5.528-21.343-12.815-23.057-6.292-1.485 5.646 7.451 9.105 15.552 12.243 6.75 2.613 12.925 5.004 12.017 8.249-.94 3.363-6.044.558-11.62-2.51-6.26-3.443-13.12-7.217-15.366-2.96-4.243 8.05 29.271 17.526 29.547 17.595 10.828 2.807 38.326 8.756 47.938-5.323Zm44.539 0c-6.914-10.136-6.423-17.742 3.063-27.22 9.486-9.476 15.008-23.335 15.008-23.335s2.06-8.052 6.76-7.311c4.7.74 8.148 12.769-1.692 20.132-9.841 7.364 1.957 12.357 5.745 5.447 3.788-6.911 14.123-24.675 19.484-28.073 5.361-3.398 9.132-1.494 7.869 5.51-.628 3.476-6.466 9.41-11.921 14.953-5.533 5.624-10.672 10.846-9.574 12.694 2.18 3.67 9.866-4.317 9.866-4.317s24.046-21.875 29.282-16.174c4.82 5.254-2.62 9.722-14.09 16.607a712.74 712.74 0 0 0-3.01 1.81c-13.125 7.94-14.14 10.035-12.28 13.038.685 1.11 5.08-1.525 10.48-4.767 9.21-5.527 21.354-12.816 23.07-6.293 1.483 5.646-7.452 9.106-15.555 12.244-6.75 2.614-12.924 5.003-12.017 8.25.94 3.363 6.04.557 11.615-2.51 6.26-3.456 13.121-7.23 15.366-2.971 4.244 8.055-29.274 17.52-29.546 17.59-10.828 2.815-38.327 8.764-47.934-5.315Z" fill="#FFD21E"/><path d="M145.615 80.61c1.361.481 2.375 1.943 3.339 3.333 1.302 1.88 2.515 3.63 4.376 2.644a12.54 12.54 0 0 0 5.173-5.13 12.488 12.488 0 0 0-.277-12.346 12.556 12.556 0 0 0-8.925-5.962 12.615 12.615 0 0 0-5.496.392 12.58 12.58 0 0 0-6.077 4.028 12.515 12.515 0 0 0-2.809 6.71 12.493 12.493 0 0 0 1.408 7.137c.862 1.616 2.776.855 4.796.05 1.584-.631 3.233-1.288 4.492-.846Zm-59.232.002c-1.362.48-2.376 1.943-3.34 3.333-1.303 1.879-2.516 3.629-4.376 2.643a12.558 12.558 0 0 1-4.177-3.577 12.505 12.505 0 0 1-2.12-10.5 12.538 12.538 0 0 1 2.464-4.909 12.576 12.576 0 0 1 4.336-3.385 12.617 12.617 0 0 1 10.75-.023 12.573 12.573 0 0 1 4.351 3.366 12.484 12.484 0 0 1 1.4 13.848c-.864 1.616-2.778.854-4.797.048-1.583-.63-3.232-1.287-4.49-.844Zm50.068 53.706c9.39-7.37 12.837-19.402 12.837-26.814 0-5.858-3.957-4.014-10.29-.89l-.356.176c-5.814 2.868-13.552 6.687-22.047 6.687-8.496 0-16.234-3.82-22.046-6.688-6.538-3.227-10.639-5.251-10.639.717 0 7.645 3.67 20.2 13.74 27.494a21.775 21.775 0 0 1 5.494-7.009 21.863 21.863 0 0 1 7.85-4.233c1.004-.298 2.04 1.426 3.1 3.19 1.023 1.704 2.067 3.445 3.132 3.445 1.132 0 2.245-1.716 3.333-3.392 1.135-1.75 2.242-3.458 3.313-3.116a21.845 21.845 0 0 1 12.58 10.433Z" fill="#32343D"/><path d="M136 134.59c-4.915 3.837-11.447 6.41-19.954 6.41-7.993 0-14.245-2.271-19.046-5.729a21.81 21.81 0 0 1 5.524-7.003 22.042 22.042 0 0 1 7.889-4.234c1.985-.587 4.087 6.632 6.264 6.632 2.33 0 4.575-7.173 6.682-6.506a21.939 21.939 0 0 1 12.63 10.43Z" fill="#FF323D" fill-rule="nonzero"/><path d="M63.71 90.637a8.248 8.248 0 0 1-2.943 1.206 8.326 8.326 0 0 1-3.19.006 8.266 8.266 0 0 1-2.946-1.193 8.082 8.082 0 0 1-2.256-2.21 7.917 7.917 0 0 1-1.22-2.89 7.856 7.856 0 0 1 .004-3.129 7.927 7.927 0 0 1 1.226-2.887 8.084 8.084 0 0 1 2.26-2.205 8.286 8.286 0 0 1 6.114-1.179 8.165 8.165 0 0 1 5.178 3.399 7.876 7.876 0 0 1 1.22 5.992 7.995 7.995 0 0 1-3.448 5.09Zm116.687 0c-.89.588-1.89.998-2.942 1.206a8.325 8.325 0 0 1-3.188.006 8.269 8.269 0 0 1-2.948-1.193 8.075 8.075 0 0 1-2.255-2.21 7.869 7.869 0 0 1 .01-8.905 8.086 8.086 0 0 1 2.26-2.206 8.287 8.287 0 0 1 6.113-1.179 8.167 8.167 0 0 1 5.178 3.399 7.876 7.876 0 0 1 1.22 5.992 7.995 7.995 0 0 1-3.448 5.09Z" fill="#FFAD03"/></g></svg> | |
| """ | |
| with gr.Blocks(css=css) as demo: | |
| gr.HTML(f""" | |
| <div class="header-container"> | |
| <div>{hf_logo_svg}</div><br /> | |
| <div class="leaderboard-title">Open LLM Leaderboard</div> | |
| </div> | |
| <div class="leaderboard-subtitle">Comparing Large Language Models in an open and reproducible way</div> | |
| """) | |
| with gr.Row(): | |
| filters = gr.CheckboxGroup( | |
| label="Quick Filters", | |
| choices=filter_choices, | |
| ) | |
| with gr.Row(): | |
| status_text = gr.HTML("<div style='text-align: right; color: #888; font-size: 0.8rem;'>Last updated: June 25, 2024 at 10:30 AM</div>") | |
| leaderboard_df = create_leaderboard_data() | |
| leaderboard_table = gr.Dataframe( | |
| value=leaderboard_df, | |
| headers=["Rank", "Type", "Model", "Average", "IFEval", "BBHI", "MATH", "GPQA", "MUJB", "MMLU-PRO", "CO_Cost"], | |
| datatype=["number", "str", "markdown", "str", "str", "str", "str", "str", "str", "str", "str"], | |
| elem_id="leaderboard-table", | |
| interactive=False, | |
| max_height=600, | |
| show_search="search", | |
| show_copy_button=True, | |
| show_fullscreen_button=True, | |
| pinned_columns=2, | |
| column_widths=["5%", "7%", "35%", "7%", "7%", "7%", "7%", "7%", "7%", "7%", "6%"] | |
| ) | |
| refresh_btn = gr.Button("Refresh Data") | |
| refresh_btn.click(fn=lambda: create_leaderboard_data(), outputs=leaderboard_table) | |
| if __name__ == "__main__": | |
| demo.launch() | |