Update app.py

app.py

@@ -9,23 +9,17 @@ import numpy as np
 import pandas as pd
 
 
-# notebook_url = "https://colab.research.google.com/drive/1RAWb22-PFNI-X1gPVzc927SGUdfr6nsR?usp=sharing"
-notebook_url = "https://colab.research.google.com/drive/1KdwokPjirkTmpO_P1WByFNFiqxWQquwH#scrollTo=o_CpbkGEbhrK"
-
 basic_component_values = [None] * 6
 leader_component_values = [None] * 5
 
-
 def make_default_md(arena_df, elo_results):
     total_votes = sum(arena_df["num_battles"]) // 2
     total_models = len(arena_df)
 
     leaderboard_md = f"""
-# 🏆 LMSYS Chatbot Arena Leaderboard
-| [Vote](https://chat.lmsys.org) | [Blog](https://lmsys.org/blog/2023-05-03-arena/) | [GitHub](https://github.com/lm-sys/FastChat) | [Paper](https://arxiv.org/abs/2306.05685) | [Dataset](https://github.com/lm-sys/FastChat/blob/main/docs/dataset_release.md) | [Twitter](https://twitter.com/lmsysorg) | [Discord](https://discord.gg/HSWAKCrnFx) |
+# NeurIPS LLM Merging Competition Leaderboard
+[Website]() | [GitHub]() | [Discord]() |
 
-LMSYS [Chatbot Arena](https://lmsys.org/blog/2023-05-03-arena/) is a crowdsourced open platform for LLM evals.
-We've collected over **500,000** human preference votes to rank LLMs with the Elo ranking system.
 """
     return leaderboard_md
 
@@ -35,11 +29,8 @@ def make_arena_leaderboard_md(arena_df):
     total_models = len(arena_df)
     space = "   "
     leaderboard_md = f"""
-Total #models: **{total_models}**.{space} Total #votes: **{"{:,}".format(total_votes)}**.{space} Last updated: April 11, 2024.
+Total #models: **{total_models}**.{space} Total #votes: **{"{:,}".format(total_votes)}**.{space} Last updated: June 1, 2024.
 
-📣 **NEW!** View leaderboard for different categories (e.g., coding, long user query)!
-
-Code to recreate leaderboard tables and plots in this [notebook]({notebook_url}). Cast your vote 🗳️ at [chat.lmsys.org](https://chat.lmsys.org)!
 """
     return leaderboard_md
 
@@ -56,14 +47,9 @@ def make_category_arena_leaderboard_md(arena_df, arena_subset_df, name="Overall"
 
 def make_full_leaderboard_md(elo_results):
     leaderboard_md = f"""
-Three benchmarks are displayed: **Arena Elo**, **MT-Bench** and **MMLU**.
-- [Chatbot Arena](https://chat.lmsys.org/?arena) - a crowdsourced, randomized battle platform. We use 500K+ user votes to compute Elo ratings.
-- [MT-Bench](https://arxiv.org/abs/2306.05685): a set of challenging multi-turn questions. We use GPT-4 to grade the model responses.
-- [MMLU](https://arxiv.org/abs/2009.03300) (5-shot): a test to measure a model's multitask accuracy on 57 tasks.
-
-💻 Code: The MT-bench scores (single-answer grading on a scale of 10) are computed by [fastchat.llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge).
-The MMLU scores are mostly computed by [InstructEval](https://github.com/declare-lab/instruct-eval).
-Higher values are better for all benchmarks. Empty cells mean not available.
+Three benchmarks are displayed: **Task 1**, **Task 2**, **Task 3**.
+
+Higher values are better for all benchmarks.
 """
     return leaderboard_md
 
@@ -334,13 +320,11 @@ cat_name_to_explanation = {
     "Exclude Short": "User Query >= 5 tokens",
 }
 
-
 def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=False):
     arena_dfs = {}
     category_elo_results = {}
     if elo_results_file is None:  # Do live update
         default_md = "Loading ..."
-        p1 = p2 = p3 = p4 = None
     else:
         with open(elo_results_file, "rb") as fin:
             elo_results = pickle.load(fin)
@@ -352,10 +336,6 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                     arena_dfs[key_to_category_name[k]] = elo_results[k]["leaderboard_table_df"]
                     category_elo_results[key_to_category_name[k]] = elo_results[k]
 
-        p1 = category_elo_results["Overall"]["win_fraction_heatmap"]
-        p2 = category_elo_results["Overall"]["battle_count_heatmap"]
-        p3 = category_elo_results["Overall"]["bootstrap_elo_rating"]
-        p4 = category_elo_results["Overall"]["average_win_rate_bar"]
         arena_df = arena_dfs["Overall"]
         default_md = make_default_md(arena_df, category_elo_results["Overall"])
 
@@ -382,8 +362,6 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                         "Rank",
                         "🤖 Model",
                         "⭐ Arena Elo",
-                        "📊 95% CI",
-                        "🗳️ Votes",
                         "Organization",
                         "License",
                         "Knowledge Cutoff",
@@ -393,55 +371,23 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                         "markdown",
                         "number",
                         "str",
-                        "number",
-                        "str",
                         "str",
                         "str",
                     ],
                     value=arena_table_vals,
                     elem_id="arena_leaderboard_dataframe",
                     height=700,
-                    column_widths=[70, 190, 110, 100, 90, 160, 150, 140],
+                    column_widths=[70, 190, 110, 160, 150, 140],
                     wrap=True,
                 )
 
                 gr.Markdown(
-                    f"""Note: we take the 95% confidence interval into account when determining a model's ranking.
-            A model is ranked higher only if its lower bound of model score is higher than the upper bound of the other model's score.
-            See Figure 3 below for visualization of the confidence intervals. More details in [notebook]({notebook_url}).
+                    f"""Note: .
             """,
                     elem_id="leaderboard_markdown"
                 )
 
-                leader_component_values[:] = [default_md, p1, p2, p3, p4]
-
-                if show_plot:
-                    more_stats_md = gr.Markdown(
-                        f"""## More Statistics for Chatbot Arena (Overall)""",
-                        elem_id="leaderboard_header_markdown"
-                    )
-                    with gr.Row():
-                        with gr.Column():
-                            gr.Markdown(
-                                "#### Figure 1: Fraction of Model A Wins for All Non-tied A vs. B Battles", elem_id="plot-title"
-                            )
-                            plot_1 = gr.Plot(p1, show_label=False, elem_id="plot-container")
-                        with gr.Column():
-                            gr.Markdown(
-                                "#### Figure 2: Battle Count for Each Combination of Models (without Ties)", elem_id="plot-title"
-                            )
-                            plot_2 = gr.Plot(p2, show_label=False)
-                    with gr.Row():
-                        with gr.Column():
-                            gr.Markdown(
-                                "#### Figure 3: Confidence Intervals on Model Strength (via Bootstrapping)", elem_id="plot-title"
-                            )
-                            plot_3 = gr.Plot(p3, show_label=False)
-                        with gr.Column():
-                            gr.Markdown(
-                                "#### Figure 4: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)", elem_id="plot-title"
-                            )
-                            plot_4 = gr.Plot(p4, show_label=False)
+                leader_component_values[:] = [default_md]
 
             with gr.Tab("Full Leaderboard", id=1):
                 md = make_full_leaderboard_md(elo_results)
@@ -450,9 +396,9 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                 gr.Dataframe(
                     headers=[
                         "🤖 Model",
-                        "⭐ Arena Elo",
-                        "📈 MT-bench",
-                        "📚 MMLU",
+                        "⭐ Task 1",
+                        "📈 Task 2",
+                        "📚 Task 3",
                         "Organization",
                         "License",
                     ],
@@ -465,8 +411,7 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                 )
         if not show_plot:
             gr.Markdown(
-                """ ## Visit our [HF space](https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard) for more analysis!
-                If you want to see more models, please help us [add them](https://github.com/lm-sys/FastChat/blob/main/docs/arena.md#how-to-add-a-new-model).
+                """ ## Submti your model [here]().
                 """,
                 elem_id="leaderboard_markdown",
             )
@@ -474,7 +419,7 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
         pass
 
     def update_leaderboard_df(arena_table_vals):
-        elo_datarame = pd.DataFrame(arena_table_vals, columns=[ "Rank", "Delta", "🤖 Model", "⭐ Arena Elo", "📊 95% CI", "🗳️ Votes", "Organization", "License", "Knowledge Cutoff"])
+        elo_datarame = pd.DataFrame(arena_table_vals, columns=[ "Rank", "🤖 Model", "⭐ Arena Elo", "Organization", "License", "Knowledge Cutoff"])
 
         # goal: color the rows based on the rank with styler
         def highlight_max(s):
@@ -484,7 +429,7 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
         def highlight_rank_max(s):
             return ["color: green; font-weight: bold" if v > 0 else "color: red; font-weight: bold" if v < 0 else "" for v in s]
         
-        return elo_datarame.style.apply(highlight_max, subset=["Rank"]).apply(highlight_rank_max, subset=["Delta"])
+        return elo_datarame.style.apply(highlight_max, subset=["Rank"])
 
     def update_leaderboard_and_plots(category):
         arena_subset_df = arena_dfs[category]
@@ -497,30 +442,24 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
             arena_values = gr.Dataframe(
                     headers=[
                         "Rank",
-                        "Delta",
                         "🤖 Model",
                         "⭐ Arena Elo",
-                        "📊 95% CI",
-                        "🗳️ Votes",
                         "Organization",
                         "License",
                         "Knowledge Cutoff",
                     ],
                     datatype=[
-                        "number",
                         "number",
                         "markdown",
                         "number",
                         "str",
-                        "number",
-                        "str",
                         "str",
                         "str",
                     ],
                     value=arena_values,
                     elem_id="arena_leaderboard_dataframe",
                     height=700,
-                    column_widths=[60, 70, 190, 110, 100, 90, 160, 150, 140],
+                    column_widths=[60, 190, 110, 160, 150, 140],
                     wrap=True,
                 )
         else:
@@ -529,8 +468,6 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                         "Rank",
                         "🤖 Model",
                         "⭐ Arena Elo",
-                        "📊 95% CI",
-                        "🗳️ Votes",
                         "Organization",
                         "License",
                         "Knowledge Cutoff",
@@ -540,28 +477,21 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                         "markdown",
                         "number",
                         "str",
-                        "number",
-                        "str",
                         "str",
                         "str",
                     ],
                     value=arena_values,
                     elem_id="arena_leaderboard_dataframe",
                     height=700,
-                    column_widths=[70, 190, 110, 100, 90, 160, 150, 140],
+                    column_widths=[70, 190, 110, 160, 150, 140],
                     wrap=True,
                 )
 
-        p1 = elo_subset_results["win_fraction_heatmap"]
-        p2 = elo_subset_results["battle_count_heatmap"]
-        p3 = elo_subset_results["bootstrap_elo_rating"]
-        p4 = elo_subset_results["average_win_rate_bar"]
-        more_stats_md = f"""## More Statistics for Chatbot Arena - {category}
-        """
+
         leaderboard_md = make_category_arena_leaderboard_md(arena_df, arena_subset_df, name=category)
-        return arena_values, p1, p2, p3, p4, more_stats_md, leaderboard_md
+        return arena_values, more_stats_md, leaderboard_md
                 
-    category_dropdown.change(update_leaderboard_and_plots, inputs=[category_dropdown], outputs=[elo_display_df, plot_1, plot_2, plot_3, plot_4, more_stats_md, category_deets])
+    category_dropdown.change(update_leaderboard_and_plots, inputs=[category_dropdown], outputs=[elo_display_df, more_stats_md, category_deets])
 
     with gr.Accordion(
         "📝 Citation",
@@ -569,22 +499,14 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
     ):
         citation_md = """
         ### Citation
-        Please cite the following paper if you find our leaderboard or dataset helpful.
-        ```
-        @misc{chiang2024chatbot,
-            title={Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference},
-            author={Wei-Lin Chiang and Lianmin Zheng and Ying Sheng and Anastasios Nikolas Angelopoulos and Tianle Li and Dacheng Li and Hao Zhang and Banghua Zhu and Michael Jordan and Joseph E. Gonzalez and Ion Stoica},
-            year={2024},
-            eprint={2403.04132},
-            archivePrefix={arXiv},
-            primaryClass={cs.AI}
-        }
+        Please cite the following paper 
+    
         """
         gr.Markdown(citation_md, elem_id="leaderboard_markdown")
         gr.Markdown(acknowledgment_md)
 
     if show_plot:
-        return [md_1, plot_1, plot_2, plot_3, plot_4]
+        return [md_1]
     return [md_1]
 
 
@@ -656,15 +578,9 @@ footer {
 
 acknowledgment_md = """
 ### Acknowledgment
-We thank [Kaggle](https://www.kaggle.com/), [MBZUAI](https://mbzuai.ac.ae/), [a16z](https://www.a16z.com/), [Together AI](https://www.together.ai/), [Anyscale](https://www.anyscale.com/), [HuggingFace](https://huggingface.co/) for their generous [sponsorship](https://lmsys.org/donations/).
+We thank []() for their generous [sponsorship]().
 
 <div class="sponsor-image-about">
-    <img src="https://storage.googleapis.com/public-arena-asset/kaggle.png" alt="Kaggle">
-    <img src="https://storage.googleapis.com/public-arena-asset/mbzuai.jpeg" alt="MBZUAI">
-    <img src="https://storage.googleapis.com/public-arena-asset/a16z.jpeg" alt="a16z">
-    <img src="https://storage.googleapis.com/public-arena-asset/together.png" alt="Together AI">
-    <img src="https://storage.googleapis.com/public-arena-asset/anyscale.png" alt="AnyScale">
-    <img src="https://storage.googleapis.com/public-arena-asset/huggingface.png" alt="HuggingFace">
 </div>
 """
 
@@ -674,7 +590,7 @@ def build_demo(elo_results_file, leaderboard_table_file):
     theme.set(button_secondary_background_fill_hover="*primary_300", 
               button_secondary_background_fill_hover_dark="*primary_700")
     with gr.Blocks(
-        title="Chatbot Arena Leaderboard",
+        title="LLM Merging Leaderboard",
         theme=theme,
         # theme = gr.themes.Base.load("theme.json"), # uncomment to use new cool theme
         css=block_css,