app.py

import hashlib
import os

import pandas as pd
import plotly.express as px
import streamlit as st
from bat import Benchmark, Config, Reporter, Tester
from datetime import datetime


st.set_page_config(
    page_title="BenchBench",
    page_icon="🏋️‍♂️",
    layout="wide",
    initial_sidebar_state="auto",
    menu_items=None,
)

# # Inject custom CSS to set the width of the sidebar
# st.markdown(
#     """
#     <style>
#         section[data-testid="stSidebar"] {
#             width: 200px !important; # Set the width to your desired value
#         }
#     </style>
#     """,
#     unsafe_allow_html=True,
# )

holistic_scenarios = [
    "Helm Lite",
    "HF OpenLLM v2",
    "OpenCompass Academic",
    "LMSys Arena",
    "Helm Classic",
    "AlphacaEval v2lc",
    "LiveBench 240725",
    "WildBench Elo LC",
]


st.markdown(
    """
    <h1 style='text-align: center; color: black;'>🏋️‍♂️ BenchBench Leaderboard 🏋️‍♂️</h1>
    """,
    unsafe_allow_html=True,
)

st.divider()

st.markdown(
    """
    The BenchBench leaderboard ranks benchmarks based on their agreement with the *Aggregate Benchmark* – a comprehensive, combined measure of existing benchmark results.
    \n
    To achive it, we scraped results from multiple benchmarks (citations below) to allow for obtaining benchmark agreement results with a wide range of benchmark using a large set of models.
    \n
    BenchBench is for you if:
    """
)

st.markdown(
    """
    - **You have a new benchmark**: Show that it agrees/disagrees with known benchmarks.
    - **You are looking for a benchmark to run/trust**: Find an efficient/private/preferble alternative.
    """
)

st.markdown(
    """
    In our work -- [Benchmark Agreement Testing Done Right](https://arxiv.org/abs/2407.13696), 
    we standardize BAT and show the importance of its configurations, notably, 
    the benchmarks we compare to, and the models we use to compare with, check it out int he sidebar.
    \n
    We show that agreements are best reporesented with the Z Score, the relative agreement of each benchmark to the Aggragate benchmark, as presented below.
    """
)

all_scenarios_for_aggragate = Benchmark()
all_scenarios_for_aggragate.load_local_catalog()
all_scenarios_for_aggragate = (
    all_scenarios_for_aggragate.df["scenario"].unique().tolist()
)

with st.sidebar:
    st.markdown("""# Configurations""")

    # with st.expander("Leaderboard configurations (defaults are great BTW)", icon="⚙️"):
    with st.form("my_form_1"):
        aggragate_scenarios = st.multiselect(
            "Aggregate Benchmark",
            all_scenarios_for_aggragate,
            holistic_scenarios,
        )

        corr_type = st.selectbox(
            label="Correlation type", options=["kendall", "pearson"], index=0
        )

        aggregate_scenario_whitelist = aggragate_scenarios
        # [
        #     scen
        #     for scen in all_scenarios_for_aggragate
        #     if scen not in aggragate_scenarios
        # ]

        model_select_strategy = st.selectbox(
            label="Model Select strategy",
            options=["random", "top_aggregate", "somewhere_aggregate"],
            index=0,
        )

        n_models_taken_list = st.slider(
            label="Minimal number of models to use",
            min_value=3,
            max_value=15,
            value=8,
        )

        n_models_taken_list = [n_models_taken_list]

        n_exps = 5

        submitted = st.form_submit_button(label="Run BAT")


with st.expander("Add your benchmarks here!", icon="🔥"):
    aggbench = Benchmark()
    aggbench.load_local_catalog()

    aggbench.add_aggregate(
        new_col_name="aggregate",
        agg_source_name="aggregate",
        scenario_whitelist=aggregate_scenario_whitelist,
        min_scenario_for_models_to_appear_in_agg=1
        if len(aggregate_scenario_whitelist) == 1
        else 3,
    )

    agg_models = (
        aggbench.df.query('scenario=="aggregate"').sample(n=10)["model"].tolist()
    )

    st.markdown(
        "Adding your benchmark is as simple as uploading a csv with the following format, one column indicates the model and the other the benchmark scores."
    )

    st.dataframe(
        pd.read_csv("assets/mybench_240901.csv"),
        use_container_width=True,
        hide_index=True,
        height=200,
    )

    st.markdown(
        "Not sure, what models you should run your benchmark on?" "\ntry these:"
    )

    st.code(agg_models)

    st.markdown("Got the data? Upload it here 👇:")

    uploaded_file = st.file_uploader("Add your benchmark as a CSV")

    my_benchmark = Benchmark()
    if uploaded_file is not None:
        st.markdown(
            "Your benchmark has been uploaded, BAT results will soon be caluclated... check out its results here: [Benchmark BAT Report Card](#benchmark-report-card)"
        )

        df = pd.read_csv(uploaded_file)

        my_benchmark.assign_df(
            df,
            data_source=f"uploaded_benchmark_{datetime.now().strftime('%y%m%d')}.csv",
            normalized_names=False,
        )

        uploaded_models = my_benchmark.df[
            my_benchmark.df["source"].str.contains("uploaded")
        ]["model"].unique()
        aggregate_models = aggbench.df[aggbench.df["source"].str.contains("aggregate")][
            "model"
        ].unique()

        # Find the intersection (overlap) of models
        overlap_models = set(aggregate_models).intersection(uploaded_models)
        if len(overlap_models) < n_models_taken_list[0]:
            st.warning(
                f"You have just {len(overlap_models)} models intersecting with the aggregate!\n"
            )

            st.info(
                f"Here are some models you could run your benchmark over:{[m for m in aggregate_models if m not in uploaded_models]}"
            )
            st.info(
                f"Model that you have and the aggragate does not: {[m for m in uploaded_models if m not in aggregate_models]}"
            )


def run_load(
    aggregate_scenario_whitelist,
    n_models_taken_list=[5],
    model_select_strategy_list=["random"],
    corr_types=["kendall"],
    n_exps=10,
    my_benchmark=Benchmark(),
    use_caching=True,
):
    # Create a hash of the inputs to generate a unique cache file for each set of inputs
    input_str = (
        str(aggregate_scenario_whitelist)
        + str(n_models_taken_list)
        + str(model_select_strategy_list)
        + str(corr_types)
        + str(n_exps)
    )

    if not my_benchmark.is_empty:
        input_str += str(
            hashlib.sha256(
                my_benchmark.df.to_csv(index=False).encode("utf-8")
            ).hexdigest()
        )

    input_hash = hashlib.md5(input_str.encode()).hexdigest()
    cache_file = f"agreements_cache_{input_hash}.csv"

    # Define the cache directory
    cache_dir = "cache"
    os.makedirs(cache_dir, exist_ok=True)
    cache_path = os.path.join(cache_dir, cache_file)

    # Check if the cache file exists
    if os.path.exists(cache_path) and use_caching:
        print("Loading cached results...")
        agreements = pd.read_csv(cache_path)
        aggregate_scores = pd.read_csv(
            cache_path.replace("agreement", "aggregate_scores")
        )
        allbench = Benchmark(
            pd.read_csv(cache_path.replace("agreement", "allbench")),
            normalized_names=True,
        )

        return agreements, aggregate_scores, allbench

    else:
        print("Cached results not found, calculating")

        allbench = Benchmark()
        allbench.load_local_catalog()

        scenarios_to_drop = ["HFv2 BBH Raw"]
        allbench.df = allbench.df.query("scenario not in @scenarios_to_drop")

        allbench.add_aggregate(
            new_col_name="aggregate",
            agg_source_name="aggregate",
            scenario_whitelist=aggregate_scenario_whitelist,
            min_scenario_for_models_to_appear_in_agg=1
            if len(aggregate_scenario_whitelist) == 1
            else len(aggregate_scenario_whitelist) // 3,
        )

        allbench.extend(my_benchmark)
        allbench.clear_repeated_scenarios()

        aggragate_scores = allbench.df.query('scenario=="aggregate"')[
            ["model", "score"]
        ].sort_values(by="score", ascending=False)

        if not my_benchmark.is_empty:
            aggragate_scores["in_uploaded"] = aggragate_scores["model"].apply(
                lambda x: x in my_benchmark.df["model"].unique()
            )

            # Get unique models for each scenario
            uploaded_models = allbench.df[
                allbench.df["source"].str.contains("uploaded")
            ]["model"].unique()
            aggregate_models = allbench.df[
                allbench.df["source"].str.contains("aggregate")
            ]["model"].unique()

            # Find the intersection (overlap) of models
            n_overlap_models = len(set(aggregate_models).intersection(uploaded_models))
            # make sure we are asking for the maximal number of models between the request benchmark and the aggregate
            n_models_taken_list = [min(n_models_taken_list[0], n_overlap_models)]

        cfg = Config(
            exp_to_run="example",
            n_models_taken_list=n_models_taken_list,
            model_select_strategy_list=model_select_strategy_list,
            corr_types=corr_types,
            n_exps=n_exps if n_models_taken_list != [0] else 1,
        )

        tester = Tester(cfg=cfg)

        agreements = tester.all_vs_all_agreement_testing(
            allbench,
            single_source_scenario="aggregate",  # olny measuring all with the aggragate
        )

        agreements.to_csv(cache_path, index=False)
        aggragate_scores.to_csv(
            cache_path.replace("agreement", "aggregate_scores"), index=False
        )
        allbench.df.to_csv(cache_path.replace("agreement", "allbench"), index=False)

    return agreements, aggragate_scores, allbench


agreements, aggragare_score_df, allbench = run_load(
    aggregate_scenario_whitelist=aggregate_scenario_whitelist,
    n_models_taken_list=n_models_taken_list,
    model_select_strategy_list=[model_select_strategy],
    corr_types=[corr_type],
    n_exps=n_exps,
    my_benchmark=my_benchmark,
)

reporter = Reporter()
z_scores = reporter.get_all_z_scores(agreements=agreements, aggragate_name="aggregate")
z_scores.drop(columns=["n_models_of_corr_with_agg"], inplace=True)

corr_name = f"{'Kendall Tau' if corr_type=='kendall' else 'Per.'} Corr."

z_scores["z_score"] = z_scores["z_score"].round(2)
z_scores["corr_with_agg"] = z_scores["corr_with_agg"].round(2)
z_scores["p_value_of_corr_with_agg"] = z_scores["p_value_of_corr_with_agg"].round(2)
# z_scores["n_models_of_corr_with_agg"] = z_scores["n_models_of_corr_with_agg"].round(1)

z_scores["date"] = z_scores["source"].apply(
    lambda x: x.split(".csv")[0].split("_")[-1]
    if "frozen" not in x
    else x.split(".csv")[0].split("_")[-2]
)

z_scores["date"] = pd.to_datetime("20" + z_scores["date"]).dt.date

z_score_name = "Relative agreement (Z Score)"

data = (
    z_scores.rename(
        columns={
            "scenario": "Benchmark",
            "z_score": z_score_name,
            "corr_with_agg": corr_name,
            "p_value_of_corr_with_agg": "p-value of Corr.",
            # "n_models_of_corr_with_agg": "# Models Used",
            "source": "Source",
            "date": "Snapshot Date",
        }
    )
    .sort_values(z_score_name, ascending=False)
    .reset_index(drop=True)
)


# Apply coloring based on 'Z' valuesz
def highlight_uploaded_benchmark(row):
    if "uploaded_benchmark" in row["Source"]:
        return ["background-color: rgba(100,100,100,0.1)"] * len(row)
    else:
        return [""] * len(row)


styled_data = (
    data.style.background_gradient(
        subset=[z_score_name],
        cmap="RdYlGn",
        vmin=-data[z_score_name].abs().max(),
        vmax=data[z_score_name].abs().max(),
    )
    .apply(highlight_uploaded_benchmark, axis=1)
    .background_gradient(
        subset=["p-value of Corr."],
        cmap="Reds",
        vmin=0.1,
        vmax=1,
    )
    .format(subset=[z_score_name, corr_name, "p-value of Corr."], formatter="{:.2}")
    .set_properties(**{"text-align": "center"})
)

cols_used = [
    "Benchmark",
    z_score_name,
    corr_name,
    "p-value of Corr.",
    "Snapshot Date",
]


st.dataframe(
    data=styled_data,
    column_order=cols_used,
    hide_index=True,
    use_container_width=True,
    height=500,
    column_config={col: {"alignment": "center"} for col in cols_used},
)


aggragare_score_df.rename(
    columns={
        "model": "Model",
        "score": "Mean Win Rate over Selected Scenarios for Aggragate",
    },
    inplace=True,
)

with st.expander(label="Aggragate Benchmark scores"):
    st.dataframe(
        data=aggragare_score_df,
        hide_index=True,
        height=500,
        use_container_width=True,
    )

with st.expander(label="Citations"):
    st.code(
        r"""

    @misc{berkeley-function-calling-leaderboard,
        title={Berkeley Function Calling Leaderboard}, 
        author={Fanjia Yan and Huanzhi Mao and Charlie Cheng-Jie Ji
        and Tianjun Zhang and Shishir G. Patil and Ion Stoica and Joseph E.
        Gonzalez},
        howpublished={\url{https://gorilla.cs.berkeley.edu/blogs/8_berkeley_function_calling_leaderboard.html}},
        year={2024},
    }

    @misc{liu2023agentbenchevaluatingllmsagents,
        title={AgentBench: Evaluating LLMs as Agents}, 
        author={Xiao Liu and Hao Yu and Hanchen Zhang and Yifan Xu and Xuanyu Lei and Hanyu Lai and Yu Gu and Hangliang Ding and Kaiwen Men and Kejuan Yang and Shudan Zhang and Xiang Deng and Aohan Zeng and Zhengxiao Du and Chenhui Zhang and Sheng Shen and Tianjun Zhang and Yu Su and Huan Sun and Minlie Huang and Yuxiao Dong and Jie Tang},
        year={2023},
        eprint={2308.03688},
        archivePrefix={arXiv},
        primaryClass={cs.AI},
        url={https://arxiv.org/abs/2308.03688}, 
    }
            
    @software{Li_AlpacaEval_An_Automatic_2023,
        author = {Li, Xuechen and Zhang, Tianyi and Dubois, Yann and Taori, Rohan and Gulrajani, Ishaan and Guestrin, Carlos and Liang, Percy and Hashimoto, Tatsunori B.},
        month = may,
        title = {{AlpacaEval: An Automatic Evaluator of Instruction-following Models}},
        year = {2023}
    }
            
    @misc{li2024crowdsourceddatahighqualitybenchmarks,
        title={From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline}, 
        author={Tianle Li and Wei-Lin Chiang and Evan Frick and Lisa Dunlap and Tianhao Wu and Banghua Zhu and Joseph E. Gonzalez and Ion Stoica},
        year={2024},
        eprint={2406.11939},
        archivePrefix={arXiv},
        primaryClass={cs.LG},
        url={https://arxiv.org/abs/2406.11939}, 
    }
    @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}
    }
    @misc{arenahard2024,
        title = {From Live Data to High-Quality Benchmarks: The Arena-Hard Pipeline},
        url = {https://lmsys.org/blog/2024-04-19-arena-hard/},
        author = {Tianle Li*, Wei-Lin Chiang*, Evan Frick, Lisa Dunlap, Banghua Zhu, Joseph E. Gonzalez, Ion Stoica},
        month = {April},
        year = {2024}
    }
            
    @misc{kim2024biggenbenchprincipledbenchmark,
        title={The BiGGen Bench: A Principled Benchmark for Fine-grained Evaluation of Language Models with Language Models}, 
        author={Seungone Kim and Juyoung Suk and Ji Yong Cho and Shayne Longpre and Chaeeun Kim and Dongkeun Yoon and Guijin Son and Yejin Cho and Sheikh Shafayat and Jinheon Baek and Sue Hyun Park and Hyeonbin Hwang and Jinkyung Jo and Hyowon Cho and Haebin Shin and Seongyun Lee and Hanseok Oh and Noah Lee and Namgyu Ho and Se June Joo and Miyoung Ko and Yoonjoo Lee and Hyungjoo Chae and Jamin Shin and Joel Jang and Seonghyeon Ye and Bill Yuchen Lin and Sean Welleck and Graham Neubig and Moontae Lee and Kyungjae Lee and Minjoon Seo},
        year={2024},
        eprint={2406.05761},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2406.05761}, 
    }
                
    @misc{liang2023holisticevaluationlanguagemodels,
        title={Holistic Evaluation of Language Models}, 
        author={Percy Liang and Rishi Bommasani and Tony Lee and Dimitris Tsipras and Dilara Soylu and Michihiro Yasunaga and Yian Zhang and Deepak Narayanan and Yuhuai Wu and Ananya Kumar and Benjamin Newman and Binhang Yuan and Bobby Yan and Ce Zhang and Christian Cosgrove and Christopher D. Manning and Christopher Ré and Diana Acosta-Navas and Drew A. Hudson and Eric Zelikman and Esin Durmus and Faisal Ladhak and Frieda Rong and Hongyu Ren and Huaxiu Yao and Jue Wang and Keshav Santhanam and Laurel Orr and Lucia Zheng and Mert Yuksekgonul and Mirac Suzgun and Nathan Kim and Neel Guha and Niladri Chatterji and Omar Khattab and Peter Henderson and Qian Huang and Ryan Chi and Sang Michael Xie and Shibani Santurkar and Surya Ganguli and Tatsunori Hashimoto and Thomas Icard and Tianyi Zhang and Vishrav Chaudhary and William Wang and Xuechen Li and Yifan Mai and Yuhui Zhang and Yuta Koreeda},
        year={2023},
        eprint={2211.09110},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2211.09110}, 
    }


    @misc{open-llm-leaderboard-v2,
    author = {Clémentine Fourrier and Nathan Habib and Alina Lozovskaya and Konrad Szafer and Thomas Wolf},
    title = {Open LLM Leaderboard v2},
    year = {2024},
    publisher = {Hugging Face},
    howpublished = "\url{https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard}",
    }

    @software{eval-harness,
        author       = {Gao, Leo and
                        Tow, Jonathan and
                        Biderman, Stella and
                        Black, Sid and
                        DiPofi, Anthony and
                        Foster, Charles and
                        Golding, Laurence and
                        Hsu, Jeffrey and
                        McDonell, Kyle and
                        Muennighoff, Niklas and
                        Phang, Jason and
                        Reynolds, Laria and
                        Tang, Eric and
                        Thite, Anish and
                        Wang, Ben and
                        Wang, Kevin and
                        Zou, Andy},
        title        = {A framework for few-shot language model evaluation},
        month        = sep,
        year         = 2021,
        publisher    = {Zenodo},
        version      = {v0.0.1},
        doi          = {10.5281/zenodo.5371628},
        url          = {https://doi.org/10.5281/zenodo.5371628},
    }

    @misc{zhou2023instructionfollowingevaluationlargelanguage,
        title={Instruction-Following Evaluation for Large Language Models},
        author={Jeffrey Zhou and Tianjian Lu and Swaroop Mishra and Siddhartha Brahma and Sujoy Basu and Yi Luan and Denny Zhou and Le Hou},
        year={2023},
        eprint={2311.07911},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2311.07911},
    }

    @misc{suzgun2022challengingbigbenchtaskschainofthought,
        title={Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them},
        author={Mirac Suzgun and Nathan Scales and Nathanael Schärli and Sebastian Gehrmann and Yi Tay and Hyung Won Chung and Aakanksha Chowdhery and Quoc V. Le and Ed H. Chi and Denny Zhou and Jason Wei},
        year={2022},
        eprint={2210.09261},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2210.09261},
    }

    @misc{hendrycks2021measuringmathematicalproblemsolving,
        title={Measuring Mathematical Problem Solving With the MATH Dataset},
        author={Dan Hendrycks and Collin Burns and Saurav Kadavath and Akul Arora and Steven Basart and Eric Tang and Dawn Song and Jacob Steinhardt},
        year={2021},
        eprint={2103.03874},
        archivePrefix={arXiv},
        primaryClass={cs.LG},
        url={https://arxiv.org/abs/2103.03874},
    }

    @misc{rein2023gpqagraduatelevelgoogleproofqa,
        title={GPQA: A Graduate-Level Google-Proof Q&A Benchmark},
        author={David Rein and Betty Li Hou and Asa Cooper Stickland and Jackson Petty and Richard Yuanzhe Pang and Julien Dirani and Julian Michael and Samuel R. Bowman},
        year={2023},
        eprint={2311.12022},
        archivePrefix={arXiv},
        primaryClass={cs.AI},
        url={https://arxiv.org/abs/2311.12022},
    }

    @misc{sprague2024musrtestinglimitschainofthought,
        title={MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning},
        author={Zayne Sprague and Xi Ye and Kaj Bostrom and Swarat Chaudhuri and Greg Durrett},
        year={2024},
        eprint={2310.16049},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2310.16049},
    }

    @misc{wang2024mmluprorobustchallengingmultitask,
        title={MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark},
        author={Yubo Wang and Xueguang Ma and Ge Zhang and Yuansheng Ni and Abhranil Chandra and Shiguang Guo and Weiming Ren and Aaran Arulraj and Xuan He and Ziyan Jiang and Tianle Li and Max Ku and Kai Wang and Alex Zhuang and Rongqi Fan and Xiang Yue and Wenhu Chen},
        year={2024},
        eprint={2406.01574},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2406.01574},
    }

    @misc{open-llm-leaderboard-v1,
        author = {Edward Beeching and Clémentine Fourrier and Nathan Habib and Sheon Han and Nathan Lambert and Nazneen Rajani and Omar Sanseviero and Lewis Tunstall and Thomas Wolf},
        title = {Open LLM Leaderboard (2023-2024)},
        year = {2023},
        publisher = {Hugging Face},
        howpublished = "\url{https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard}"
    }
            
    @misc{open-llm-leaderboard,
        author = {Edward Beeching and Clémentine Fourrier and Nathan Habib and Sheon Han and Nathan Lambert and Nazneen Rajani and Omar Sanseviero and Lewis Tunstall and Thomas Wolf},
        title = {Open LLM Leaderboard},
        year = {2023},
        publisher = {Hugging Face},
        howpublished = "\url{https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard}"
    }

    @misc{waldis2024holmesbenchmarklinguisticcompetence,
        title={Holmes: Benchmark the Linguistic Competence of Language Models}, 
        author={Andreas Waldis and Yotam Perlitz and Leshem Choshen and Yufang Hou and Iryna Gurevych},
        year={2024},
        eprint={2404.18923},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2404.18923}, 
    }       

    @article{livebench,
        author    = {White, Colin and Dooley, Samuel and Roberts, Manley and Pal, Arka and Feuer, Ben and Jain, Siddhartha and Shwartz-Ziv, Ravid and Jain, Neel and Saifullah, Khalid and Naidu, Siddartha and Hegde, Chinmay and LeCun, Yann and Goldstein, Tom and Neiswanger, Willie and Goldblum, Micah},
        title     = {LiveBench: A Challenging, Contamination-Free LLM Benchmark},
        url       = {arXiv preprint arXiv:2406.19314},
        year      = {2024},
    }  
            
    @article{ni2024mixeval,
        title={MixEval: Deriving Wisdom of the Crowd from LLM Benchmark Mixtures},
        author={Ni, Jinjie and Xue, Fuzhao and Yue, Xiang and Deng, Yuntian and Shah, Mahir and Jain, Kabir and Neubig, Graham and You, Yang},
        journal={arXiv preprint arXiv:2406.06565},
        year={2024}
    }
            
    @misc{wang2024mmluprorobustchallengingmultitask,
        title={MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark}, 
        author={Yubo Wang and Xueguang Ma and Ge Zhang and Yuansheng Ni and Abhranil Chandra and Shiguang Guo and Weiming Ren and Aaran Arulraj and Xuan He and Ziyan Jiang and Tianle Li and Max Ku and Kai Wang and Alex Zhuang and Rongqi Fan and Xiang Yue and Wenhu Chen},
        year={2024},
        eprint={2406.01574},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2406.01574}, 
    }
                
    @misc{zheng2023judgingllmasajudgemtbenchchatbot,
        title={Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena}, 
        author={Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric P. Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica},
        year={2023},
        eprint={2306.05685},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2306.05685}, 
    }
            
    @misc{2023opencompass,
        title={OpenCompass: A Universal Evaluation Platform for Foundation Models},
        author={OpenCompass Contributors},
        howpublished = {\url{https://github.com/open-compass/opencompass}},
        year={2023}
    }
            
    @misc{qin2023toolllm,
        title={ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs}, 
        author={Yujia Qin and Shihao Liang and Yining Ye and Kunlun Zhu and Lan Yan and Yaxi Lu and Yankai Lin and Xin Cong and Xiangru Tang and Bill Qian and Sihan Zhao and Runchu Tian and Ruobing Xie and Jie Zhou and Mark Gerstein and Dahai Li and Zhiyuan Liu and Maosong Sun},
        year={2023},
        eprint={2307.16789},
        archivePrefix={arXiv},
        primaryClass={cs.AI}
    }
                
    @misc{qin2023tool,
        title={Tool Learning with Foundation Models}, 
        author={Yujia Qin and Shengding Hu and Yankai Lin and Weize Chen and Ning Ding and Ganqu Cui and Zheni Zeng and Yufei Huang and Chaojun Xiao and Chi Han and Yi Ren Fung and Yusheng Su and Huadong Wang and Cheng Qian and Runchu Tian and Kunlun Zhu and Shihao Liang and Xingyu Shen and Bokai Xu and Zhen Zhang and Yining Ye and Bowen Li and Ziwei Tang and Jing Yi and Yuzhang Zhu and Zhenning Dai and Lan Yan and Xin Cong and Yaxi Lu and Weilin Zhao and Yuxiang Huang and Junxi Yan and Xu Han and Xian Sun and Dahai Li and Jason Phang and Cheng Yang and Tongshuang Wu and Heng Ji and Zhiyuan Liu and Maosong Sun},
        year={2023},
        eprint={2304.08354},
        archivePrefix={arXiv},
        primaryClass={cs.CL}
    }
                
    @misc{guo2024stabletoolbench,
        title={StableToolBench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models},
        author={Guo, Zhicheng and Cheng, Sijie and Wang, Hao and Liang, Shihao and Qin, Yujia and Li, Peng and Liu, Zhiyuan and Sun, Maosong and Liu, Yang},
        year={2024},
        eprint={2403.07714},
        archivePrefix={arXiv},
        primaryClass={cs.CL}
    }
            
    @article{yuchen2024wildbench,
        title={WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild},
        author={Yuchen Lin, Bill and Deng, Yuntian and Chandu, Khyathi and Brahman, Faeze and Ravichander, Abhilasha and Pyatkin, Valentina and Dziri, Nouha and Le Bras, Ronan and Choi, Yejin},
        journal={arXiv e-prints},
        pages={arXiv--2406},
        year={2024}
    }
    
    """
    )


st.subheader("Benchmark Report Card")


benchmarks = allbench.df["scenario"].unique().tolist()
index_to_use = 1
if not my_benchmark.is_empty:
    index_to_use = benchmarks.index(my_benchmark.df["scenario"].unique()[0])

plotted_scenario = st.selectbox(
    "Choose Benchmark to plot",
    benchmarks,
    index=index_to_use,
)

col1, col2, col3 = st.columns(3)
cur_data = data.query(f"Benchmark=='{plotted_scenario}'")
col1.metric("Relative agreement", cur_data["Relative agreement (Z Score)"])
col2.metric(corr_name, cur_data[corr_name])
col3.metric("p-value of Corr.", cur_data["p-value of Corr."])

cur_df = allbench.df.query(f'scenario=="aggregate" or scenario=="{plotted_scenario}"')

# Filter models that are present in both scenarios
models_in_both = cur_df.groupby("model")["scenario"].nunique().eq(2).index

# Pivot the DataFrame to have scenarios as columns
df_pivot = cur_df[cur_df["model"].isin(models_in_both)].pivot(
    index="model", columns="scenario", values="score"
)

# Create the scatter plot using Plotly Express
fig = px.scatter(
    df_pivot,
    x=df_pivot.columns[0],
    y=df_pivot.columns[1],
    trendline="ols",
    labels={
        df_pivot.columns[0]: df_pivot.columns[0],
        df_pivot.columns[1]: df_pivot.columns[1],
    },
    hover_name=df_pivot.index,
    title="Model Scores Comparison between Scenarios",
)
st.plotly_chart(fig, use_container_width=True)

st.markdown(
    "BenchBench-Leaderboard complements our study, where we analyzed over 40 prominent benchmarks and introduced standardized practices to enhance the robustness and validity of benchmark evaluations through the [BenchBench Python package](#). "
    "The BenchBench-Leaderboard serves as a dynamic platform for benchmark comparison and is an essential tool for researchers and practitioners in the language model field aiming to select and utilize benchmarks effectively. "
)

st.subheader("How did we get the Z Scores?", divider=True)

st.write(r"""
            Section 3.1 in our work shows how using a single reference benchmark drastically hurts the roubustness and validity of BAT.
            To remedy this, we propose to test benchmark agreement with an aggragate benchmark and compare the agreement to other benchmarks.
            We recommend to perform this comparison using the [Z score](https://en.wikipedia.org/wiki/Standard_score) and demonstrate obtaining it to a benchmark of your selection.
            In the follwing way: $z_i=(x_i-\mu_{i...N}) / \sigma_{i...N}$ where $x_i$ is the agreement of the $i$th benchmark to the aggragate and $\mu_{i...N}$,$\sigma_{i...N}$ are the 
            mean and standard deviation of the agreements of the other benchmarks to the aggragate.
            """)


fig = px.histogram(
    data.query("Benchmark!=@plotted_scenario"), x=corr_name, nbins=len(data) - 1
)
# Add a vertical line at a specific x-coordinate
# Replace 'x_value' with the actual value where you want the line
x_value = 0.5  # Example value, adjust as necessary
fig.add_vline(
    x=data.query("Benchmark==@plotted_scenario")[corr_name].iloc[0],
    line_dash="dash",
    line_color="red",
)
# Update layout to add a title
fig.update_layout(
    title="Histogram of Correlation Values",  # Change the title text as needed
    title_x=0.3,  # Centers the title
    title_font=dict(size=20, family="CMU"),  # Customize font if needed
)

# # Plot!
st.plotly_chart(fig, use_container_width=True)

st.subheader("Why should you use the BenchBench Leaderboard?")

st.markdown(
    """

    Current practices in Benchmark Agreement Testing (BAT) often suffer from a lack of standardization and transparency, which can lead to inconsistent results and diminished trust in benchmark evaluations. Several key issues are prevalent in the field:

    """
)

st.markdown(
    """
    - **Lack of Standard Methodologies:** Unlike other scientific procedures that follow rigorous methodologies, BAT lacks uniform procedures across different studies. Researchers often employ varied criteria for selecting benchmarks and models for comparison, which leads to results that cannot be easily compared or replicated. This variation undermines the reliability of conclusions drawn from BAT and makes it difficult for other researchers to build on existing work.
    """
)

st.image(
    "images/motivation.png",
    caption="Conclusions depend on the models considered. Kendall-tau correlations between the LMSys Arena benchmark and three other benchmarks: BBH, MMLU, and Alpaca v2. Each group of bars represents the correlation for different sets of top models, specifically the top 5, top 10, and top 15 (overlapping) models (according to the Arena). The results indicate that the degree of agreement between benchmarks varies with the number of top models considered, highlighting that different selections of models can lead to varying conclusions about benchmark agreement.",
    use_column_width=True,
)

st.markdown(
    """
    - **Arbitrary Selection of Reference Benchmarks:** One of the most critical decisions in BAT is the choice of reference benchmarks. Currently, this choice is often arbitrary and lacks a clear rationale, influenced by availability or personal preference rather than strategic alignment with the benchmark’s purpose. This can skew the results significantly, as different benchmarks may not be equally representative or relevant to the models being tested.
    """
)
st.markdown(
    """
    - **Inadequate Model Representation:** BAT frequently relies on a limited subset of models, which may not comprehensively represent the diversity of architectures and training paradigms in modern language models. This selective representation can lead to biased agreement scores that favor certain types of models over others, failing to provide a holistic view of model performance across different benchmarks.
    """
)

st.image(
    "images/pointplot_granularity_matters.png",
    caption="Correlations increase with number of models. Mean correlation (y) between each benchmark (lines) and the rest, given different numbers of models. The Blue and Orange lines are the average of all benchmark pair correlations with models sampled randomly (orange) or in contiguous sets (blue). The shaded lines represents adjacent sampling for the different benchmarks.",
    use_column_width=True,
)

st.markdown(
    """
    - **Overemphasis on Correlation Metrics:** Current BAT practices tend to over-rely on correlation metrics without adequately considering their limitations and the context of their application. While these metrics can provide useful insights, they are often treated as definitive evidence of agreement without acknowledging that high correlation does not necessarily imply conceptual alignment between benchmarks.
    """
)

st.markdown(
    """
    To address these issues, there is a critical need for a more structured approach to BAT that includes clear guidelines for benchmark and model selection, a broader consideration of agreement metrics, and an acknowledgment of the evolving nature of technology in this space. By reforming BAT practices, the research community can improve the reliability and utility of benchmarks as tools for evaluating and advancing language models.
    """
)


st.image(
    "images/ablations.png",
    caption="Our recommendations substantially reduce the variance of BAT. Ablation analysis for each BAT recommendation separately and their combinations.",
    use_column_width=True,
)


st.header("The BenchBench package")

st.markdown("""
### Overview

The BAT package is designed to facilitate benchmark agreement testing for NLP models. It allows users to easily compare multiple models against various benchmarks and generate comprehensive reports on their agreement.

### Installation

To install the BAT package, you can use pip:

```
pip install bat-package
```

### Usage Example

Below is a step-by-step example of how to use the BAT package to perform agreement testing.

#### Step 1: Configuration

First, set up the configuration for the tests:

```python
import pandas as pd
from bat import Tester, Config, Benchmark, Reporter
from bat.utils import get_holistic_benchmark

cfg = Config(
    exp_to_run="example",
    n_models_taken_list=[0],
    model_select_strategy_list=["random"],
    n_exps=10
)
```

#### Step 2: Fetch Model Names

Fetch the names of the reference models to be used for scoring:

```python
tester = Tester(cfg=cfg)
models_for_benchmark_scoring = tester.fetch_reference_models_names(
    reference_benchmark=get_holistic_benchmark(), n_models=20
)
print(models_for_benchmark_scoring)
```

#### Step 3: Load and Prepare Benchmark

Load a new benchmark and add an aggregate column:

```python
newbench_name = "fakebench"
newbench = Benchmark(
    pd.read_csv(f"src/bat/assets/{newbench_name}.csv"),
    data_source=newbench_name,
)
newbench.add_aggregate(new_col_name=f"{newbench_name}_mwr")
```

#### Step 4: Agreement Testing

Perform all-vs-all agreement testing on the new benchmark:

```python
newbench_agreements = tester.all_vs_all_agreement_testing(newbench)
reporter = Reporter()
reporter.draw_agreements(newbench_agreements)
```

#### Step 5: Extend and Clean Benchmark

Extend the new benchmark with holistic data and clear repeated scenarios:

```python
allbench = newbench.extend(get_holistic_benchmark())
allbench.clear_repeated_scenarios(source_to_keep=newbench_name)
```

#### Step 6: Comprehensive Agreement Testing

Perform comprehensive agreement testing and visualize:

```python
all_agreements = tester.all_vs_all_agreement_testing(allbench)
reporter.draw_agreements(all_agreements)
```
""")