"""A gradio app that renders a static leaderboard. This is used for Hugging Face Space."""
import ast
import argparse
import glob
import pickle
import gradio as gr
import numpy as np
import pandas as pd
# notebook_url = "https://colab.research.google.com/drive/1RAWb22-PFNI-X1gPVzc927SGUdfr6nsR?usp=sharing"
notebook_url = "https://colab.research.google.com/drive/1KdwokPjirkTmpO_P1WByFNFiqxWQquwH#scrollTo=o_CpbkGEbhrK"
basic_component_values = [None] * 6
leader_component_values = [None] * 5
def make_default_md():
leaderboard_md = f"""
# 🏆 BabilongLeaderboard
| [GitHub](https://github.com/lm-sys/FastChat) | [Paper](https://arxiv.org/abs/2306.05685) | [Dataset](https://github.com/lm-sys/FastChat/blob/main/docs/dataset_release.md) |
"""
return leaderboard_md
def make_arena_leaderboard_md():
total_models = 'UNK'
leaderboard_md = f"""Total #models: **{total_models}**. Last updated: Feb 28, 2024."""
return leaderboard_md
def make_full_leaderboard_md(elo_results):
leaderboard_md = f"""
Three benchmarks are displayed: **Arena Elo**, **MT-Bench** and **MMLU**.
- [Chatbot Arena](https://chat.lmsys.org/?arena) - a crowdsourced, randomized battle platform. We use 200K+ user votes to compute Elo ratings.
- [MT-Bench](https://arxiv.org/abs/2306.05685): a set of challenging multi-turn questions. We use GPT-4 to grade the model responses.
- [MMLU](https://arxiv.org/abs/2009.03300) (5-shot): a test to measure a model's multitask accuracy on 57 tasks.
💻 Code: The MT-bench scores (single-answer grading on a scale of 10) are computed by [fastchat.llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge).
The MMLU scores are mostly computed by [InstructEval](https://github.com/declare-lab/instruct-eval).
Higher values are better for all benchmarks. Empty cells mean not available.
"""
return leaderboard_md
def make_leaderboard_md_live(elo_results):
leaderboard_md = f"""
# Leaderboard
Last updated: {elo_results["last_updated_datetime"]}
{elo_results["leaderboard_table"]}
"""
return leaderboard_md
def update_elo_components(max_num_files, elo_results_file):
log_files = get_log_files(max_num_files)
# Leaderboard
if elo_results_file is None: # Do live update
battles = clean_battle_data(log_files)
elo_results = report_elo_analysis_results(battles)
leader_component_values[0] = make_leaderboard_md_live(elo_results)
leader_component_values[1] = elo_results["win_fraction_heatmap"]
leader_component_values[2] = elo_results["battle_count_heatmap"]
leader_component_values[3] = elo_results["bootstrap_elo_rating"]
leader_component_values[4] = elo_results["average_win_rate_bar"]
# Basic stats
basic_stats = report_basic_stats(log_files)
md0 = f"Last updated: {basic_stats['last_updated_datetime']}"
md1 = "### Action Histogram\n"
md1 += basic_stats["action_hist_md"] + "\n"
md2 = "### Anony. Vote Histogram\n"
md2 += basic_stats["anony_vote_hist_md"] + "\n"
md3 = "### Model Call Histogram\n"
md3 += basic_stats["model_hist_md"] + "\n"
md4 = "### Model Call (Last 24 Hours)\n"
md4 += basic_stats["num_chats_last_24_hours"] + "\n"
basic_component_values[0] = md0
basic_component_values[1] = basic_stats["chat_dates_bar"]
basic_component_values[2] = md1
basic_component_values[3] = md2
basic_component_values[4] = md3
basic_component_values[5] = md4
def update_worker(max_num_files, interval, elo_results_file):
while True:
tic = time.time()
update_elo_components(max_num_files, elo_results_file)
durtaion = time.time() - tic
print(f"update duration: {durtaion:.2f} s")
time.sleep(max(interval - durtaion, 0))
def load_demo(url_params, request: gr.Request):
logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")
return basic_component_values + leader_component_values
def model_hyperlink(model_name, link):
return f'{model_name}'
def load_leaderboard_table_csv(filename, add_hyperlink=True):
lines = open(filename).readlines()
heads = [v.strip() for v in lines[0].split(",")]
rows = []
for i in range(1, len(lines)):
row = [v.strip() for v in lines[i].split(",")]
for j in range(len(heads)):
item = {}
for h, v in zip(heads, row):
if h == "Arena Elo rating":
if v != "-":
v = int(ast.literal_eval(v))
else:
v = np.nan
elif h == "MMLU":
if v != "-":
v = round(ast.literal_eval(v) * 100, 1)
else:
v = np.nan
elif h == "MT-bench (win rate %)":
if v != "-":
v = round(ast.literal_eval(v[:-1]), 1)
else:
v = np.nan
elif h == "MT-bench (score)":
if v != "-":
v = round(ast.literal_eval(v), 2)
else:
v = np.nan
item[h] = v
if add_hyperlink:
item["Model"] = model_hyperlink(item["Model"], item["Link"])
rows.append(item)
return rows
def build_basic_stats_tab():
empty = "Loading ..."
basic_component_values[:] = [empty, None, empty, empty, empty, empty]
md0 = gr.Markdown(empty)
gr.Markdown("#### Figure 1: Number of model calls and votes")
plot_1 = gr.Plot(show_label=False)
with gr.Row():
with gr.Column():
md1 = gr.Markdown(empty)
with gr.Column():
md2 = gr.Markdown(empty)
with gr.Row():
with gr.Column():
md3 = gr.Markdown(empty)
with gr.Column():
md4 = gr.Markdown(empty)
return [md0, plot_1, md1, md2, md3, md4]
def get_full_table(arena_df, model_table_df):
values = []
for i in range(len(model_table_df)):
row = []
model_key = model_table_df.iloc[i]["key"]
model_name = model_table_df.iloc[i]["Model"]
# model display name
row.append(model_name)
if model_key in arena_df.index:
idx = arena_df.index.get_loc(model_key)
row.append(round(arena_df.iloc[idx]["rating"]))
else:
row.append(np.nan)
row.append(model_table_df.iloc[i]["MT-bench (score)"])
row.append(model_table_df.iloc[i]["MMLU"])
# Organization
row.append(model_table_df.iloc[i]["Organization"])
# license
row.append(model_table_df.iloc[i]["License"])
values.append(row)
values.sort(key=lambda x: -x[1] if not np.isnan(x[1]) else 1e9)
return values
def build_leaderboard_tab():
default_md = make_default_md()
md_1 = gr.Markdown(default_md, elem_id="leaderboard_markdown")
with gr.Tabs() as tabs:
# arena table
with gr.Tab("Arena Elo", id=0):
md = make_arena_leaderboard_md()
gr.Markdown(md, elem_id="leaderboard_markdown")
gr.Dataframe(
headers=[
"Rank",
"🤖 Model",
"⭐ Arena Elo",
"📊 95% CI",
"🗳️ Votes",
"Organization",
"License",
"Knowledge Cutoff",
],
datatype=[
"str",
"markdown",
"number",
"str",
"number",
"str",
"str",
"str",
],
# value=arena_table_vals,
elem_id="arena_leaderboard_dataframe",
height=700,
column_widths=[50, 200, 120, 100, 100, 150, 150, 100],
wrap=True,
)
return [md_1]
block_css = """
#notice_markdown {
font-size: 104%
}
#notice_markdown th {
display: none;
}
#notice_markdown td {
padding-top: 6px;
padding-bottom: 6px;
}
#leaderboard_markdown {
font-size: 104%
}
#leaderboard_markdown td {
padding-top: 6px;
padding-bottom: 6px;
}
#leaderboard_dataframe td {
line-height: 0.1em;
}
footer {
display:none !important
}
.image-container {
display: flex;
align-items: center;
padding: 1px;
}
.image-container img {
margin: 0 30px;
height: 20px;
max-height: 100%;
width: auto;
max-width: 20%;
}
"""
acknowledgment_md = """
### Acknowledgment
"""
def build_demo(leaderboard_table_file):
text_size = gr.themes.sizes.text_lg
with gr.Blocks(
title="Babilong leaderboard",
theme=gr.themes.Base(text_size=text_size),
css=block_css,
) as demo:
leader_components = build_leaderboard_tab(
leaderboard_table_file, show_plot=True
)
return demo
if __name__ == "__main__":
elo_result_files = glob.glob("elo_results_*.pkl")
leaderboard_table_files = glob.glob("leaderboard_table_*.csv")
# leaderboard_table_files.sort(key=lambda x: int(x[18:-4]))
# leaderboard_table_file = leaderboard_table_files[-1]
demo = build_demo(None, None)
demo.launch(share=True)
