Upload app.py

Code cleanup and added comparison table to compare models.

app.py

@@ -1,3 +1,4 @@
+# Importing necessary libraries
 import re
 import streamlit as st
 import requests
@@ -6,11 +7,35 @@ 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
+from functools import cache
+
+
 
+# Function to get model info from Hugging Face API using caching
+@cache
+def cached_model_info(api, model):
+    try:
+        return api.model_info(repo_id=str(model))
+    except (RepositoryNotFoundError, RevisionNotFoundError):
+        return None
 
+# Function to get model info from DataFrame and update it with likes and tags
+@st.cache
+def get_model_info(df):
+    api = HfApi()
 
+    for index, row in df.iterrows():
+        model_info = cached_model_info(api, row['Model'].strip())
+        if model_info:
+            df.loc[index, 'Likes'] = model_info.likes
+            df.loc[index, 'Tags'] = ', '.join(model_info.tags)
+        else:
+            df.loc[index, 'Likes'] = -1
+            df.loc[index, 'Tags'] = ''
+    return df
+
+# Function to convert markdown table to DataFrame and extract Hugging Face URLs
 def convert_markdown_table_to_dataframe(md_content):
     """
     Converts markdown table to Pandas DataFrame, handling special characters and links,
@@ -59,8 +84,7 @@ def get_model_info(df):
 
     return df
 
-
-
+# Function to create bar chart for a given category
 def create_bar_chart(df, category):
     """Create and display a bar chart for a given category."""
     st.write(f"### {category} Scores")
@@ -73,7 +97,7 @@ def create_bar_chart(df, category):
         x=sorted_df[category],
         y=sorted_df['Model'],
         orientation='h',
-        marker=dict(color=sorted_df[category], colorscale='Inferno')
+        marker=dict(color=sorted_df[category], colorscale='Spectral')  # You can change 'Viridis' to another color scale
     ))
 
     # Update layout for better readability
@@ -82,17 +106,17 @@ def create_bar_chart(df, category):
     )
 
     # Adjust the height of the chart based on the number of rows in the DataFrame
-    st.plotly_chart(fig, use_container_width=True, height=35)
-
-# Example usage:
-# create_bar_chart(your_dataframe, 'Your_Category')
-
+    st.plotly_chart(fig, use_container_width=True, height=len(df) * 35)
 
+# Main function to run the Streamlit app
 def main():
+    # Set page configuration and title
     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.")
+
+    # Create tabs for leaderboard and about section
     content = create_yall()
     tab1, tab2 = st.tabs(["🏆 Leaderboard", "📝 About"])
 
@@ -104,21 +128,19 @@ def main():
 
                 # Display dataframe
                 full_df = convert_markdown_table_to_dataframe(content)
+
                 for col in score_columns:
                     # Corrected use of pd.to_numeric
                     full_df[col] = pd.to_numeric(full_df[col].str.strip(), errors='coerce')
+
                 full_df = get_model_info(full_df)
                 full_df['Tags'] = full_df['Tags'].fillna('')
                 df = pd.DataFrame(columns=full_df.columns)
 
-                # Toggles
-                col1, col2, col3 = st.columns(3)
-                with col1:
-                    show_phi = st.checkbox("Phi (2.8B)", value=True)
-                with col2:
-                    show_mistral = st.checkbox("Mistral (7B)", value=True)
-                with col3:
-                    show_other = st.checkbox("Other", value=True)
+                # Toggles for filtering by tags
+                show_phi = st.checkbox("Phi (2.8B)", value=True)
+                show_mistral = st.checkbox("Mistral (7B)", value=True)
+                show_other = st.checkbox("Other", value=True)
 
                 # Create a DataFrame based on selected filters
                 dfs_to_concat = []
@@ -135,9 +157,6 @@ def main():
                 if dfs_to_concat:
                     df = pd.concat(dfs_to_concat, ignore_index=True)
 
-                # Sort values
-                df = df.sort_values(by='Average', ascending=False)
-
                 # Add a search bar
                 search_query = st.text_input("Search models", "")
 
@@ -158,9 +177,11 @@ def main():
                         "URL": st.column_config.LinkColumn("URL"),
                     },
                     hide_index=True,
-                    height=int(len(df) * 36.2),
+                    height=len(df) * 37,
                 )
-
+                selected_models = st.multiselect('Select models to compare', df['Model'].unique())
+                comparison_df = df[df['Model'].isin(selected_models)]
+                st.dataframe(comparison_df)
                 # Add a button to export data to CSV
                 if st.button("Export to CSV"):
                     # Export the DataFrame to CSV
@@ -203,27 +224,28 @@ def main():
     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 and [CultriX](https://huggingface.co/CultriX) for the CSV export and search bar.
+
+            A special thanks to [gblazex](https://huggingface.co/gblazex) for providing many evaluations.
         ''')
-        
+
+# Run the main function if this script is run directly
 if __name__ == "__main__":
     main()