app.py

import re
import streamlit as st
import requests
import pandas as pd
from io import StringIO
import plotly.graph_objs as go
from huggingface_hub import HfApi
from huggingface_hub.utils import RepositoryNotFoundError, RevisionNotFoundError

from yall import create_yall


def convert_markdown_table_to_dataframe(md_content):
    """
    Converts markdown table to Pandas DataFrame, handling special characters and links,
    extracts Hugging Face URLs, and adds them to a new column.
    """
    # Remove leading and trailing | characters
    cleaned_content = re.sub(r'\|\s*$', '', re.sub(r'^\|\s*', '', md_content, flags=re.MULTILINE), flags=re.MULTILINE)

    # Create DataFrame from cleaned content
    df = pd.read_csv(StringIO(cleaned_content), sep="\|", engine='python')

    # Remove the first row after the header
    df = df.drop(0, axis=0)

    # Strip whitespace from column names
    df.columns = df.columns.str.strip()

    # Extract Hugging Face URLs and add them to a new column
    model_link_pattern = r'\[(.*?)\]\((.*?)\)\s*\[.*?\]\(.*?\)'
    df['URL'] = df['Model'].apply(lambda x: re.search(model_link_pattern, x).group(2) if re.search(model_link_pattern, x) else None)

    # Clean Model column to have only the model link text
    df['Model'] = df['Model'].apply(lambda x: re.sub(model_link_pattern, r'\1', x))

    return df

@st.cache_data 
def get_model_info(df):
    api = HfApi()
    
    # Initialize new columns for likes and tags
    df['likes'] = None
    df['tags'] = None

    # Iterate through DataFrame rows
    for index, row in df.iterrows():
        model = row['Model'].strip()
        try:
            model_info = api.model_info(repo_id=str(model))
            df.loc[index, 'Likes'] = model_info.likes
            df.loc[index, 'Tags'] = ', '.join(model_info.tags)

        except (RepositoryNotFoundError, RevisionNotFoundError):
            df.loc[index, 'Likes'] = -1
            df.loc[index, 'Tags'] = ''

    return df

    
def create_bar_chart(df, category):
    """Create and display a bar chart for a given category."""
    st.write(f"### {category} Scores")

    # Sort the DataFrame based on the category score
    sorted_df = df[['Model', category]].sort_values(by=category, ascending=True)

    # Create the bar chart with color gradient
    fig = go.Figure(go.Bar(
        x=sorted_df[category],
        y=sorted_df['Model'],
        orientation='h',
        marker=dict(color=sorted_df[category], colorscale='Inferno')
    ))

    # Update layout for better readability
    fig.update_layout(
        margin=dict(l=20, r=20, t=20, b=20)
    )

    st.plotly_chart(fig, use_container_width=True)

    
def main():
    st.set_page_config(page_title="YALL - Yet Another LLM Leaderboard", layout="wide")

    st.title("🏆 YALL - Yet Another LLM Leaderboard")
    st.markdown("Leaderboard made with [🧐 LLM AutoEval](https://github.com/mlabonne/llm-autoeval) using [Nous](https://huggingface.co/NousResearch) benchmark suite. It's a collection of my own evaluations.")

    content = create_yall()
    tab1, tab2 = st.tabs(["🏆 Leaderboard", "📝 About"])

    # Leaderboard tab
    with tab1:
        if content:
            try:
                score_columns = ['Average', 'AGIEval', 'GPT4All', 'TruthfulQA', 'Bigbench']
    
                # Display dataframe
                full_df = convert_markdown_table_to_dataframe(content)
                for col in score_columns:
                    full_df[col] = full_df.to_numeric(full_df[col].str.strip(), errors='coerce')
                full_df = get_model_info(full_df)
                df = pd.DataFrame(columns=full_df.columns)
                
                # Toggles
                col1, col2 = st.columns(2)
                with col1:
                    show_phi = st.toggle("Phi (2.8B)", value=True)
                with col2:
                    show_mistral = st.toggle("Mistral (7B)", value=True)

                # Apply filters based on toggles
                if show_phi:
                    df = df.append(full_df[full_df['tags'].str.lower().str.contains('phi-msft')], ignore_index=True)
                if show_mistral:
                    df = df.append(full_df[full_df['tags'].str.lower().str.contains('mistral')], ignore_index=True)

                # Sort values
                df = df.sort_values(by='Average', ascending=False)

                # Display the DataFrame
                st.dataframe(df[['Model'] + score_columns + ['Likes']], use_container_width=True)

                # Full-width plot for the first category
                create_bar_chart(df, score_columns[0])
    
                # Next two plots in two columns
                col1, col2 = st.columns(2)
                with col1:
                    create_bar_chart(df, score_columns[1])
                with col2:
                    create_bar_chart(df, score_columns[2])
    
                # Last two plots in two columns
                col3, col4 = st.columns(2)
                with col3:
                    create_bar_chart(df, score_columns[3])
                with col4:
                    create_bar_chart(df, score_columns[4])
    
            except Exception as e:
                st.error("An error occurred while processing the markdown table.")
                st.error(str(e))
        else:
            st.error("Failed to download the content from the URL provided.")

     # About tab
    with tab2:
        st.markdown('''
            ### Nous benchmark suite

            Popularized by [Teknium](https://huggingface.co/teknium) and [NousResearch](https://huggingface.co/NousResearch), this benchmark suite aggregates four benchmarks:
            
            * [**AGIEval**](https://arxiv.org/abs/2304.06364) (0-shot): `agieval_aqua_rat,agieval_logiqa_en,agieval_lsat_ar,agieval_lsat_lr,agieval_lsat_rc,agieval_sat_en,agieval_sat_en_without_passage,agieval_sat_math`
            * **GPT4ALL** (0-shot): `hellaswag,openbookqa,winogrande,arc_easy,arc_challenge,boolq,piqa`
            * [**TruthfulQA**](https://arxiv.org/abs/2109.07958) (0-shot): `truthfulqa_mc`
            * [**Bigbench**](https://arxiv.org/abs/2206.04615) (0-shot): `bigbench_causal_judgement,bigbench_date_understanding,bigbench_disambiguation_qa,bigbench_geometric_shapes,bigbench_logical_deduction_five_objects,bigbench_logical_deduction_seven_objects,bigbench_logical_deduction_three_objects,bigbench_movie_recommendation,bigbench_navigate,bigbench_reasoning_about_colored_objects,bigbench_ruin_names,bigbench_salient_translation_error_detection,bigbench_snarks,bigbench_sports_understanding,bigbench_temporal_sequences,bigbench_tracking_shuffled_objects_five_objects,bigbench_tracking_shuffled_objects_seven_objects,bigbench_tracking_shuffled_objects_three_objects`

            ### Reproducibility

            You can easily reproduce these results using [🧐 LLM AutoEval](https://github.com/mlabonne/llm-autoeval/tree/master), a colab notebook that automates the evaluation process (benchmark: `nous`). This will upload the results to GitHub as gists. You can find the entire table with the links to the detailed results [here](https://gist.github.com/mlabonne/90294929a2dbcb8877f9696f28105fdf).

            ### Clone this space

            You can create your own leaderboard with your LLM AutoEval results on GitHub Gist. You just need to clone this space and specify two variables:

            * Change the `gist_id` in [yall.py](https://huggingface.co/spaces/mlabonne/Yet_Another_LLM_Leaderboard/blob/main/yall.py#L126).
            * Create "New Secret" in Settings > Variables and secrets (name: "github", value: [your GitHub token](https://github.com/settings/tokens))

            A special thanks to [gblazex](https://huggingface.co/gblazex) for providing many evaluations.
        ''')

if __name__ == "__main__":
    main()