import streamlit as st import pandas as pd from huggingface_hub import HfApi from huggingface_hub.utils import RepositoryNotFoundError, RevisionNotFoundError from itertools import combinations import re from functools import cache from io import StringIO from yall import create_yall import plotly.graph_objs as go from huggingface_hub import ModelCard def calculate_pages(df, items_per_page): return -(-len(df) // items_per_page) # Equivalent to math.ceil(len(df) / items_per_page) # Function to get model info from Hugging Face API using caching @st.cache_data 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_data 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, 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 calculate_highest_combined_score(data, column): # score_columns = ['Average', 'AGIEval', 'GPT4All', 'TruthfulQA', 'Bigbench'] # # Ensure the column exists and has numeric data # if column not in data.columns or not pd.api.types.is_numeric_dtype(data[column]): # return column, {} # scores = data[column].dropna().tolist() # models = data['Model'].tolist() # top_combinations = {r: [] for r in range(2, 5)} # for r in range(2, 5): # for combination in combinations(zip(scores, models), r): # combined_score = sum(score for score, _ in combination) # top_combinations[r].append((combined_score, tuple(model for _, model in combination))) # top_combinations[r].sort(key=lambda x: x[0], reverse=True) # top_combinations[r] = top_combinations[r][:5] # return column, top_combinations ## Modified function to display the results of the highest combined scores using st.dataframe #def display_highest_combined_scores(data): # score_columns = ['Average', 'AGIEval', 'GPT4All', 'TruthfulQA', 'Bigbench'] # with st.spinner('Calculating highest combined scores...'): # results = [calculate_highest_combined_score(data, col) for col in score_columns] # for column, top_combinations in results: # st.subheader(f"Top Combinations for {column}") # for r, combinations in top_combinations.items(): # # Prepare data for DataFrame # rows = [{'Score': score, 'Models': ', '.join(combination)} for score, combination in combinations] # df = pd.DataFrame(rows) # # # Display using st.dataframe # st.markdown(f"**Number of Models: {r}**") # st.dataframe(df, height=150) # Adjust height as necessary # 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") # Sort the DataFrame based on the category score sorted_df = df[['Model', category]].sort_values(by=category, ascending=True) # Create the bar chart with a color gradient (using 'Viridis' color scale as an example) fig = go.Figure(go.Bar( x=sorted_df[category], y=sorted_df['Model'], orientation='h', marker=dict(color=sorted_df[category], colorscale='Spectral') # You can change 'Viridis' to another color scale )) # Update layout for better readability fig.update_layout( margin=dict(l=20, r=20, t=20, b=20) ) # Adjust the height of the chart based on the number of rows in the DataFrame st.plotly_chart(fig, use_container_width=True, height=len(df) * 35) def fetch_merge_configs(df): # Sort the data by the second column (assuming the column name is 'Average') df_sorted = df.sort_values(by='Average', ascending=False) # Open the file in append mode with open('/tmp/configurations.txt', 'a') as file: # Get model cards for the top 20 entries for index, row in df_sorted.head(20).iterrows(): model_name = row['Model'].rstrip() card = ModelCard.load(model_name) file.write(f'Model Name: {model_name}\n') file.write(f'Scores: {row["Average"]}\n') # Assuming 'Average' is the benchmark score file.write(f'AGIEval: {row["AGIEval"]}\n') file.write(f'GPT4All: {row["GPT4All"]}\n') file.write(f'TruthfulQA: {row["TruthfulQA"]}\n') file.write(f'Bigbench: {row["Bigbench"]}\n') file.write(f'Model Card: {card}\n') # Open the second file in read mode with open('/tmp/configurations.txt', 'r') as file: # Read the content content = file.read() # Find all text between 'yaml' and '```' matches = re.findall(r'yaml(.*?)```', content, re.DOTALL) # Open the file 'configurations2.txt' in write mode with open('/tmp/configurations2.txt', 'w') as file: # Write the matches to the file for row, match in zip(df_sorted[['Model', 'Average', 'AGIEval', 'GPT4All', 'TruthfulQA', 'Bigbench']].head(20).values, matches): file.write(f'Model Name: {row[0]}\n') file.write(f'Scores: {row[1]}\n') file.write(f'AGIEval: {row[2]}\n') file.write(f'GPT4All: {row[3]}\n') file.write(f'TruthfulQA: {row[4]}\n') file.write(f'Bigbench: {row[5]}\n') file.write('yaml' + match + '```\n') # 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"]) # 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: # 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 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 = [] if show_phi: dfs_to_concat.append(full_df[full_df['Tags'].str.lower().str.contains('phi,|phi-msft,')]) if show_mistral: dfs_to_concat.append(full_df[full_df['Tags'].str.lower().str.contains('mistral,')]) if show_other: other_df = full_df[~full_df['Tags'].str.lower().str.contains('phi,|phi-msft,|mistral,')] dfs_to_concat.append(other_df) # Concatenate the DataFrames if dfs_to_concat: df = pd.concat(dfs_to_concat, ignore_index=True) # Add a search bar search_query = st.text_input("Search models", "") # Filter the DataFrame based on the search query if search_query: df = df[df['Model'].str.contains(search_query, case=False)] # Add a selectbox for page selection items_per_page = 50 pages = calculate_pages(df, items_per_page) page = st.selectbox("Page", list(range(1, pages + 1))) # Sort the DataFrame by 'Average' column in descending order df = df.sort_values(by='Average', ascending=False) # Slice the DataFrame based on the selected page start = (page - 1) * items_per_page end = start + items_per_page df = df[start:end] # Display the filtered DataFrame or the entire leaderboard st.dataframe( df[['Model'] + score_columns + ['Likes', 'URL']], use_container_width=True, column_config={ "Likes": st.column_config.NumberColumn( "Likes", help="Number of likes on Hugging Face", format="%d ❤️", ), "URL": st.column_config.LinkColumn("URL"), }, hide_index=True, 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 csv_data = df.to_csv(index=False) # Create a link to download the CSV file st.download_button( label="Download CSV", data=csv_data, file_name="leaderboard.csv", key="download-csv", help="Click to download the CSV file", ) if st.button("Fetch Merge-Configs"): # Call the function with the current DataFrame configurations, matches, csv_data = fetch_merge_configs(full_df) # Assuming full_df is your DataFrame # You can then display the configurations or matches as needed, or write them to a file # For example, displaying the configurations: for config in configurations: st.text(f"Model Name: {config['Model Name']}\nScores: {config['Scores']}\nAGIEval: {config['AGIEval']}\nGPT4All: {config['GPT4All']}\nTruthfulQA: {config['TruthfulQA']}\nBigbench: {config['Bigbench']}\nModel Card: {config['Model Card']}\n\n") # Convert the list of dictionaries to a DataFrame configurations_df = pd.DataFrame(configurations) # Convert the DataFrame to a CSV string configurations_csv = configurations_df.to_csv(index=False) st.download_button( label="Download Configurations", data=configurations_csv, file_name="configurations.csv", key="download-csv", help="Click to download the CSV file", ) # 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]) # display_highest_combined_scores(full_df) # Call to display the calculated scores 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('''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. # Bonus: Workflow for Automating Model Evaluation and Selection ## Step 1. Export CSV Data from Another-LLM-LeaderBoards Go to our [Another-LLM-LeaderBoards](https://leaderboards.example.com) and click the export csv data button. Save it to `/tmp/models.csv`. ## Step 2: Examine CSV Data Run a script for extracting model names, benchmark scores, and model page link from the CSV data. ```python import re from huggingface_hub import ModelCard import pandas as pd # Load the CSV data df = pd.read_csv('/tmp/models.csv') # Sort the data by the second column (assuming the column name is 'Average') df_sorted = df.sort_values(by='Average', ascending=False) # Open the file in append mode with open('/tmp/configurations.txt', 'a') as file: # Get model cards for the top 20 entries for index, row in df_sorted.head(20).iterrows(): model_name = row['Model'].rstrip() card = ModelCard.load(model_name) file.write(f'Model Name: {model_name}\n') file.write(f'Scores: {row["Average"]}\n') # Assuming 'Average' is the benchmark score file.write(f'AGIEval: {row["AGIEval"]}\n') file.write(f'GPT4All: {row["GPT4All"]}\n') file.write(f'TruthfulQA: {row["TruthfulQA"]}\n') file.write(f'Bigbench: {row["Bigbench"]}\n') file.write(f'Model Card: {card}\n') ``` ## Step 3: Feed the Discovered Models, Scores and Configurations to LLM-client (shell-gpt) Run your local LLM-client by feeding it all the discovered merged models, their benchmark scores and if found the configurations used to merge them. Provide it with an instruction similar to this: ```bash cat /tmp/configurations2.txt | sgpt --chat config "Based on the merged models that are provided here, along with their respective benchmark achievements and the configurations used in merging them, your task is to come up with a new configuration for a new merged model that will outperform all others. In your thought process, argue and reflect on your own choices to improve your thinking process and outcome" ``` ## Step 4: (Optional) Reflect on Initial Configuration Suggested by Chat-GPT If you wanted to get particularly naughty, you could add a step like this where you make Chat-GPT rethink and reflect on the configuration it initially comes up with based on the information you gave it. ```bash for i in $(seq 1 3); do echo "$i" && sgpt --chat config "Repeat the process from before and again reflect and improve on your suggested configuration"; sleep 20; done ``` ## Step 5: Wait for Chat-GPT to give you a LeaderBoard-topping merge configuration Wait for Chat-GPT to provide a new merge configuration. ## Step 6: Enter the Configuration in Automergekit NoteBook Fire up your automergekit NoteBook and enter in the configuration that was just so generously provided to you by Chat-GPT. ## Step 7: Evaluate the New Merge using auto-llm-eval notebook Fire up your auto-llm-eval notebook to see if the merge that Chat-GPT came up with is actually making any sense and performing well. ## Step 8: Repeat the Process Repeat this process for a few times every day, learning from each new model created. ## Step 9: Rank the New Number One Model Rank the new number one model and top your own LeaderBoard: (Model: CultriX/MergeCeption-7B-v3) ![image.png](https://cdn-uploads.huggingface.co/production/uploads/6495d5a915d8ef6f01bc75eb/mFV3Ou469fk6ivj1XrD9d.png) ## Step 10: Automate the Process with Cronjob Create a cronjob that automates this process 5 times every day, only to then learn from the models that it has created in order to create even better ones and I'm telling you that you better prepare yourself for some non-neglectable increases in benchmark scores for the near future. Cheers, CultriX ''') # Run the main function if this script is run directly if __name__ == "__main__": main()