# app.py import json import streamlit as st import glob import os from datetime import datetime #st.set_page_config(layout="wide") st.title('Meta Open LLM leaderboard') directories = os.listdir("./data") #data_dir = directories[0] def format_dir_date(data_dir): # Extracting date and time information from the path parsed_date = datetime.strptime(data_dir, "%Y%m%d_%H%M") # Formatting the parsed date return parsed_date.strftime("%b %d, %Y %H:%M") data_dir = st.selectbox( 'Select different Date', directories, format_func=format_dir_date, index=len(directories)-1 ) captions_map = { "hg_average_to_agentbench_compare.png": "HF to AgentBench compare", "hg_average_to_opencompass_compare.png": "HF to OpenCompass</a> compare", "hg_average_to_mt_bench_compare.png": "HF to MT-Bench compare", "hg_average_to_mosaic_compare.png": "HF to MosaicML compare", "hg_average_to_alpacaeval_compare.png": "HF to AlpacaEval compare" } st.write("Generated on: <b>" + format_dir_date(data_dir) + "</b>", unsafe_allow_html=True) st.divider() data_path = './data/' + data_dir imgs = glob.glob(os.path.join(data_path, '*.png')) # Extracting images that start with "hf_llm_diagram" hf_llm_diagrams = [img for img in imgs if 'hf_llm_diagram' in os.path.basename(img)] # Getting the remaining images remaining_imgs = [img for img in imgs if 'hf_llm_diagram' not in os.path.basename(img)] def print_model_list(file_name, st, split_into_two=False): file_path = file_name[:-4] + '.json' # Read the list from the JSON file with open(file_path, 'r') as file: model_id_list_loaded = json.load(file) midpoint = len(model_id_list_loaded) // 2 + (len(model_id_list_loaded) % 2) # Calculate the midpoint # Split the list into two parts left_list = model_id_list_loaded[:midpoint] right_list = model_id_list_loaded[midpoint:] # Generate HTML for the left column left_html = "" for model_id in left_list: model_id_trunc = model_id if len(model_id) <= 35 else '...' + model_id[-35:] left_html += f'<li><a href="https://huggingface.co/{model_id}">{model_id_trunc}</a></li>' # Generate HTML for the right column right_html = "" for model_id in right_list: model_id_trunc = model_id if len(model_id) <= 35 else '...' + model_id[-35:] right_html += f'<li><a href="https://huggingface.co/{model_id}">{model_id_trunc}</a></li>' final_html = "" if(split_into_two): final_html = "<ul>" final_html += left_html final_html += "</ul>" cols = st.columns(2) cols[0].write(final_html, unsafe_allow_html=True) final_html = "<ul>" final_html += right_html final_html += "</ul>" cols[1].write(final_html, unsafe_allow_html=True) else: final_html = "<ul>" final_html += left_html final_html += right_html final_html += "</ul>" st.write(final_html, unsafe_allow_html=True) st.write("HuggingFace Open LLM leaderboard by Model Size") st.image(hf_llm_diagrams[0],use_column_width="auto") print_model_list(hf_llm_diagrams[0],st,True) cols = st.columns(2) cols[0].image(hf_llm_diagrams[1],caption="Other or commercially permissive licenses only", use_column_width="auto") print_model_list(hf_llm_diagrams[1],cols[0]) cols[1].image(hf_llm_diagrams[2],caption="Commercially permissive license only", use_column_width="auto") print_model_list(hf_llm_diagrams[2],cols[1]) st.divider() st.write("HuggingFace and Other Leaderboards: A Comparative Model Evaluation") st.caption("Only models evaluated on both leaderboards are included.") cols = st.columns(2) for i, img in enumerate(remaining_imgs): # Extract the filename from the full image path filename = os.path.basename(img) # Get the caption from the captions_map dictionary caption = captions_map.get(filename, "") # If no caption is found, it will default to an empty string # Display the image with the caption cols[i % 2].image(img, caption=caption, width=None) st.write( """ <p>Leaderboards tracked:</p> <ul> <li><a href="https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard">Hugging Face Open LLM</a></li> <li><a href="https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard">MT-Bench</a> GPT4 judged evaluation of models</li> <li><a href="https://tatsu-lab.github.io/alpaca_eval/">AlpacaEval</a> GPT4 judged evaluation of models</li> <li><a href="https://www.mosaicml.com/llm-evaluation">MosaicML</a> Balanced set of static benchmarks</li> <li><a href="https://opencompass.org.cn/leaderboard-llm">OpenCompass</a> Balanced set of static benchmarks</li> <li><a href="https://llmbench.ai/data">AgentBench</a> Benchmark evaluating Agent abilities</li> </ul> """, unsafe_allow_html=True ) st.divider() st.write("TruthfulQA at 10% for HuggingFace Open LLM leaderboard by Model Size") st.image(hf_llm_diagrams[3],use_column_width="auto") print_model_list(hf_llm_diagrams[3],st,True) st.divider() st.subheader('About') st.write('This meta leaderboard is built and maintained by Felix Zaslavskiy. For feedback, correction, suggestions please reach out on X at <a href="https://twitter.com/FZaslavskiy" >@FZaslavskiy</a> or here via community discussions.', unsafe_allow_html=True)