app.py

import streamlit as st
import pandas as pd
import numpy as np
from html import escape


st.set_page_config(layout="wide")

column_config = {
    "Downloads": st.column_config.NumberColumn(
        "Downloads", format="%d 📥"
    ),
    "Likes": st.column_config.NumberColumn(
        "Likes", format="%d ❤️"
    ),
    "Hugging Face URL": st.column_config.LinkColumn("Hugging Face URL", display_text="Open"),
    "Arxiv URL": st.column_config.LinkColumn("Arxiv URL", display_text="Open"),
    "PapersWithCode URL": st.column_config.LinkColumn("PapersWithCode URL", display_text="Open")
}


@st.cache_data
def load_data():
    file_path = 'HuggingFaceBenchmarkDatasetsWithTags - Copy of HuggingFaceBenchmarkDatasetsWithTags (1).csv'
    data = pd.read_csv(file_path, na_values=['NA', ''])

    data['Created At'] = pd.to_datetime(data['Created At'], errors='coerce')
    data['Last Modified'] = pd.to_datetime(data['Last Modified'], errors='coerce')

    numeric_cols = ['Downloads', 'Likes', 'Total Examples', 'Dataset Size (bytes)']
    for col in numeric_cols:
        data[col] = pd.to_numeric(data[col], errors='coerce')

    data.replace("", np.nan, inplace=True)

    data = data.drop(columns=['Card Data', 'Model Card README'])
    return data

def escape_html(val):
    return escape(val) if isinstance(val, str) else val
    
df = load_data()

st.title('Bench1k: LLM Benchmarks & Evals Database')
st.subheader('Explore 1,327+ benchmarks. By default, sorted by # of downloads.')
st.write("Use the sidebar to apply filters.")

search_query = st.text_input("Search benchmarks by keyword")


task_ids = df['Task IDs'].dropna().unique().tolist()
selected_task_id = st.sidebar.multiselect('Filter by Task IDs', task_ids)

task_categories = df['Task Categories'].dropna().unique().tolist()
selected_task_category = st.sidebar.multiselect('Filter by Task Categories', task_categories)

licenses = df['Licenses'].dropna().unique().tolist()
selected_license = st.sidebar.multiselect('Filter by License', licenses)

min_likes, max_likes = int(df['Likes'].min(skipna=True)), int(df['Likes'].max(skipna=True))
selected_likes = st.sidebar.slider('Filter by Likes', min_likes, max_likes, (min_likes, max_likes))

min_size, max_size = df['Dataset Size (bytes)'].min(skipna=True), df['Dataset Size (bytes)'].max(skipna=True)
selected_size = st.sidebar.slider('Filter by Dataset Size (bytes)', min_size, max_size, (min_size, max_size))

min_examples, max_examples = df['Total Examples'].min(skipna=True), df['Total Examples'].max(skipna=True)
selected_examples = st.sidebar.slider('Filter by Total Examples', min_examples, max_examples, (min_examples, max_examples))

filtered_df = df

if search_query:
    search_cols = df.select_dtypes(include=[object]).columns
    filtered_df = filtered_df[filtered_df[search_cols].apply(lambda x: x.str.contains(search_query, case=False, na=False)).any(axis=1)]
    
if selected_task_id:
    mask_task_id = filtered_df['Task IDs'].apply(lambda x: any(task_id.strip() in str(x).split(',') for task_id in selected_task_id))
    filtered_df = filtered_df[mask_task_id]

if selected_task_category:
    mask_task_category = filtered_df['Task Categories'].apply(lambda x: any(category.strip() in str(x).split(',') for category in selected_task_category))
    filtered_df = filtered_df[mask_task_category]

if selected_license:
    filtered_df = filtered_df[filtered_df['Licenses'].isin(selected_license)]

if selected_likes:
    filtered_df = filtered_df[filtered_df['Likes'].between(selected_likes[0], selected_likes[1])]

if selected_size:
    filtered_df = filtered_df[filtered_df['Dataset Size (bytes)'].between(selected_size[0], selected_size[1])]

if selected_examples:
    filtered_df = filtered_df[filtered_df['Total Examples'].between(selected_examples[0], selected_examples[1])]

def clean_html_sensitive_content(val):
    if isinstance(val, str):
        val = ''.join(e for e in val if e.isalnum() or e in [' ', '-', '_'])
    return val

filtered_df['Task IDs'] = filtered_df['Task IDs'].apply(clean_html_sensitive_content)
filtered_df['Task Categories'] = filtered_df['Task Categories'].apply(clean_html_sensitive_content)

st.dataframe(filtered_df, column_config=column_config, hide_index=True)

st.sidebar.info("Use the filters above to explore different aspects of the benchmark datasets.")


@st.cache_data
def convert_df(df):
    return df.to_csv().encode('utf-8')

csv = convert_df(df)

st.download_button(
    label="Download database as CSV",
    data=csv,
    file_name='bench1k_database_full.csv',
    mime='text/csv',
)