app.py

"""A gradio app that renders a static leaderboard. This is used for Hugging Face Space.""" 
import argparse  
import gradio as gr
import numpy as np
import pandas as pd
import gradio as gr
import pandas as pd 
import json 
from constants import BANNER, INTRODUCTION_TEXT, CITATION_TEXT, METRICS_TAB_TEXT, DIR_OUTPUT_REQUESTS  

LAST_UPDATED = "Feb 28th 2024"

css = """
.markdown-text{font-size: 15pt}
.markdown-text-small{font-size: 13pt}
th {
  text-align: center;
}
td {
  font-size: 15px; /* Adjust the font size as needed */
  text-align: center;
}
#od-benchmark-tab-table-button{
    font-size: 15pt;
    font-weight: bold;
}

"""

column_names = {
    "model": "Model",
    "Overall": "All 🎯",    
    "Turn 1": "Turn 1️⃣", 
    "Turn 2": "Turn 2️⃣", 
    }

model_info = {
    "gpt-4": {"hf_name": "https://platform.openai.com/", "pretty_name": "gpt-4"}, 
    "gpt-3.5-turbo": {"hf_name": "https://platform.openai.com/", "pretty_name": "gpt-3.5-turbo"}, 
    "Llama-2-70b-hf": {"hf_name": "meta-llama/Llama-2-70b-hf", "pretty_name": "Llama-2-70B"}, 
    "Llama-2-13b-hf": {"hf_name": "meta-llama/Llama-2-13b-hf", "pretty_name": "Llama-2-13B"}, 
    "Llama-2-7b-hf": {"hf_name": "meta-llama/Llama-2-7b-hf", "pretty_name": "Llama-2-7B"}, 
    "Mixtral-8x7B-v0.1": {"hf_name": "mistralai/Mixtral-8x7B-v0.1", "pretty_name": "Mixtral-8x7B"}, 
    "Mistral-7b-v0.1": {"hf_name": "mistralai/Mistral-7B-v0.1", "pretty_name": "Mistral-7B v0.1"}, 
    "Mistral-7b-v0.2": {"hf_name": "alpindale/Mistral-7B-v0.2-hf", "pretty_name": "Mistral-7B v0.2"}, 
    "Yi-34B": {"hf_name": "01-ai/Yi-34B", "pretty_name": "Yi-34B"}, 
    "Yi-6B": {"hf_name": "01-ai/Yi-6B", "pretty_name": "Yi-6B"}, 
    "gemma-7b": {"hf_name": "google/gemma-7b", "pretty_name": "Gemma-7B"}, 
    "gemma-2b": {"hf_name": "google/gemma-2b", "pretty_name": "Gemma-2B"}, 
    "phi-2": {"hf_name": "microsoft/phi-2", "pretty_name": "Phi-2 @hf"}, 
    "olmo": {"hf_name": "allenai/OLMo-7B", "pretty_name": "OLMo-7B @hf"},  
    "phi-2-vllm": {"hf_name": "microsoft/phi-2", "pretty_name": "Phi-2 (2.7B)"}, 
    "olmo-7b-vllm": {"hf_name": "allenai/OLMo-7B", "pretty_name": "OLMo-7B"},  
    "falcon-7b": {"hf_name": "microsoft/falcon-7b", "pretty_name": "Falcon-7B"},
    "mpt-7b": {"hf_name": "mosaicml/mpt-7b", "pretty_name": "MPT-7B"},
    "amber": {"hf_name": "LLM360/Amber", "pretty_name": "Amber (7B)"},
    "dbrx": {"hf_name": "databricks/dbrx-base", "pretty_name": "DBRX-base"},
} 


def formatter(x):
    x = round(x, 2)
    return x

def make_clickable_model(model_name, model_info):
    if model_info[model_name]['hf_name'].startswith("http"):
        link = model_info[model_name]['hf_name']
    else:
        link = f"https://huggingface.co/{model_info[model_name]['hf_name']}"
    if model_name.startswith("gpt"):
        return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted; background-color: lightgray;">{model_info[model_name]["pretty_name"]}</a>'
    else:
        return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_info[model_name]["pretty_name"]}</a>'


def build_demo(original_df, full_df, TYPES):
    with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
        # gr.HTML(BANNER, elem_id="banner")
        gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")

        with gr.Tabs(elem_classes="tab-buttons") as tabs:
            with gr.TabItem("🏅 Leaderboard", elem_id="od-benchmark-tab-table", id=0):
                leaderboard_table = gr.components.Dataframe(
                    value=original_df,
                    datatype=TYPES,
                    height=1000,
                    wrap=False,
                    elem_id="leaderboard-table",
                    interactive=False,
                    visible=True,
                    min_width=60,
                    )

            with gr.TabItem("🐑 URIAL + 🤗 OpenLLM", elem_id="od-benchmark-tab-table", id=1):
                gr.Markdown("### More results from the awesome 🤗 [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) ", elem_classes="markdown-text")
                leaderboard_table_full = gr.components.Dataframe(
                    value=full_df,
                    datatype=TYPES,
                    height=1000,
                    wrap=False,
                    elem_id="leaderboard-table-full",
                    interactive=False,
                    visible=True,
                    min_width=60,
                    ) 

        gr.Markdown(f"Last updated on **{LAST_UPDATED}**", elem_classes="markdown-text-small")
        
        with gr.Row():
            with gr.Accordion("📙 Citation", open=False):
                gr.Textbox(
                    value=CITATION_TEXT, lines=18,
                    label="Copy the BibTeX to cite URIAL and MT-Bench",
                    elem_id="citation-button",
                    show_copy_button=True)
                # ).style(show_copy_button=True)

    return demo

if __name__ == "__main__": 

    parser = argparse.ArgumentParser()
    parser.add_argument("--share", action="store_true")
    parser.add_argument("--result_file", help="Path to results table", default="leaderboard_data.jsonl")
    args = parser.parse_args()

    all_model_hf_ids = {v["hf_name"]: k  for k, v in model_info.items()}

    # Load Open LLM Leaderboard 
    with open("open-llm-leaderboard.json") as f:
        open_llm_leaderbaord = json.load(f)
    full_leaderboard = {}
    for item in open_llm_leaderbaord:
        if item["Model"] in all_model_hf_ids:
            # print(item["Model"])
            # print(item["Average \u2b06\ufe0f"])
            full_bench_item = {}
            # full_bench_item["hf_name"] = item["Model"]
            full_bench_item["model_name"] = all_model_hf_ids[item["Model"]]
            tasks = ["HellaSwag", "ARC", "Winogrande", "TruthfulQA", "MMLU", "GSM8K"]
            for task in tasks:
                full_bench_item[task] = item[task]
            full_bench_item["HF_AVG"] = item["Average \u2b06\ufe0f"]
            full_leaderboard[all_model_hf_ids[item["Model"]]] = full_bench_item
    # Load URIAL Leaderboard
    with open("leaderboard_data.jsonl") as f:
        for line in f:
            item = json.loads(line)  
            if item["model"] in full_leaderboard:
                full_leaderboard[item["model"]]["URIAL_AVG"] = item["Overall"]   
 

    # Process the URIAL Benchmark Tab 
    original_df = pd.read_json(args.result_file, lines=True)

    print(original_df.columns)

    for col in original_df.columns:
        if col == "model":
            original_df[col] = original_df[col].apply(lambda x: x.replace(x, make_clickable_model(x, model_info)))
        else: 
            original_df[col] = original_df[col].apply(formatter) # For numerical values
    # Define the first column explicitly, add 'Overall' as the second column, and then append the rest excluding 'Overall'
    new_order = [original_df.columns[0], 'Overall'] + [col for col in original_df.columns if col not in [original_df.columns[0], 'Overall']]
    # Reorder the DataFrame columns using the new order
    reordered_df = original_df[new_order]
    reordered_df.sort_values(by='Overall', inplace=True, ascending=False)
    reordered_df.rename(columns=column_names, inplace=True)

    # Process the Full Benchmark Tab
    full_df = pd.DataFrame(full_leaderboard).T
    full_df = full_df.reset_index()
    full_df.rename(columns={"index": "model"}, inplace=True)
    full_df = full_df[["model",  "URIAL_AVG", "HF_AVG", "HellaSwag", "ARC", "Winogrande", "TruthfulQA", "MMLU", "GSM8K"]]
    full_df.sort_values(by='URIAL_AVG', inplace=True, ascending=False)
    full_df["model"] = full_df["model"].apply(lambda x: make_clickable_model(x, model_info))
    full_df.rename(columns=column_names, inplace=True)
    # apply formatter to numerical columns
    for col in full_df.columns:
        if col not in ["Model"]:
            full_df[col] = full_df[col].apply(formatter) # For numerical values
    # COLS = [c.name for c in fields(AutoEvalColumn)]
    # TYPES = [c.type for c in fields(AutoEvalColumn)]

    TYPES = ["markdown", "number"]
    demo = build_demo(reordered_df, full_df, TYPES)
    demo.launch(share=args.share)