andrewrreed's picture
andrewrreed HF staff
fix due to vision leaderboard addtion upstream
ec01232
raw
history blame
12.6 kB
import os
import pickle
import pandas as pd
import gradio as gr
import plotly.express as px
from datetime import datetime
from huggingface_hub import HfApi
from apscheduler.schedulers.background import BackgroundScheduler
from utils import (
KEY_TO_CATEGORY_NAME,
CAT_NAME_TO_EXPLANATION,
download_latest_data_from_space,
get_constants,
update_release_date_mapping,
format_data,
get_trendlines,
find_crossover_point,
)
###################
### Initialize scheduler
###################
def restart_space():
HfApi(token=os.getenv("HF_TOKEN", None)).restart_space(
repo_id="andrewrreed/closed-vs-open-arena-elo"
)
print(f"Space restarted on {datetime.now()}")
# restart the space every day at 9am
scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "cron", day_of_week="mon-sun", hour=7, minute=0)
scheduler.start()
###################
### Load Data
###################
# gather ELO data
latest_elo_file_local = download_latest_data_from_space(
repo_id="lmsys/chatbot-arena-leaderboard", file_type="pkl"
)
with open(latest_elo_file_local, "rb") as fin:
elo_results = pickle.load(fin)
# TO-DO: need to also include vision
elo_results = elo_results["text"]
arena_dfs = {}
for k in KEY_TO_CATEGORY_NAME.keys():
if k not in elo_results:
continue
arena_dfs[KEY_TO_CATEGORY_NAME[k]] = elo_results[k]["leaderboard_table_df"]
# gather open llm leaderboard data
latest_leaderboard_file_local = download_latest_data_from_space(
repo_id="lmsys/chatbot-arena-leaderboard", file_type="csv"
)
leaderboard_df = pd.read_csv(latest_leaderboard_file_local)
# load release date mapping data
release_date_mapping = pd.read_json("release_date_mapping.json", orient="records")
###################
### Prepare Data
###################
# update release date mapping with new models
# check for new models in ELO data
new_model_keys_to_add = [
model
for model in arena_dfs["Overall"].index.to_list()
if model not in release_date_mapping["key"].to_list()
]
if new_model_keys_to_add:
release_date_mapping = update_release_date_mapping(
new_model_keys_to_add, leaderboard_df, release_date_mapping
)
# merge leaderboard data with ELO data
merged_dfs = {}
for k, v in arena_dfs.items():
merged_dfs[k] = (
pd.merge(arena_dfs[k], leaderboard_df, left_index=True, right_on="key")
.sort_values("rating", ascending=False)
.reset_index(drop=True)
)
# add release dates into the merged data
for k, v in merged_dfs.items():
merged_dfs[k] = pd.merge(
merged_dfs[k], release_date_mapping[["key", "Release Date"]], on="key"
)
# format dataframes
merged_dfs = {k: format_data(v) for k, v in merged_dfs.items()}
# get constants
min_elo_score, max_elo_score, upper_models_per_month = get_constants(merged_dfs)
date_updated = elo_results["full"]["last_updated_datetime"].split(" ")[0]
orgs = merged_dfs["Overall"].Organization.unique().tolist()
###################
### Build and Plot Data
###################
def get_data_split(dfs, set_name):
df = dfs[set_name].copy(deep=True)
return df.reset_index(drop=True)
def clean_df_for_display(df):
df = df.loc[
:,
[
"Model",
"rating",
"MMLU",
"MT-bench (score)",
"Release Date",
"Organization",
"License",
"Link",
],
].rename(columns={"rating": "ELO Score", "MT-bench (score)": "MT-Bench"})
df["Release Date"] = df["Release Date"].astype(str)
df.sort_values("ELO Score", ascending=False, inplace=True)
df.reset_index(drop=True, inplace=True)
return df
def filter_df(min_score, max_models_per_month, set_selector, org_selector):
df = get_data_split(merged_dfs, set_name=set_selector)
# filter data
filtered_df = df[
(df["rating"] >= min_score) & (df["Organization"].isin(org_selector))
]
filtered_df = (
filtered_df.groupby(["Month-Year", "License"], group_keys=False)
.apply(lambda x: x.nlargest(max_models_per_month, "rating"))
.reset_index(drop=True)
)
return filtered_df
def build_plot(toggle_annotations, filtered_df):
# construct plot
custom_colors = {"Open": "#ff7f0e", "Proprietary": "#1f77b4"}
fig = px.scatter(
filtered_df,
x="Release Date",
y="rating",
color="License",
hover_name="Model",
hover_data=["Organization", "License", "Link"],
trendline="ols",
title=f"Open vs Proprietary LLMs by LMSYS Arena ELO Score<br>(as of {date_updated})",
labels={"rating": "Arena ELO", "Release Date": "Release Date"},
height=700,
template="plotly_dark",
color_discrete_map=custom_colors,
)
fig.update_layout(
plot_bgcolor="rgba(0,0,0,0)", # Set background color to transparent
paper_bgcolor="rgba(0,0,0,0)", # Set paper (plot) background color to transparent
title={"x": 0.5},
)
fig.update_traces(marker=dict(size=10, opacity=0.6))
# calculate days until crossover
trend1, trend2 = get_trendlines(fig)
crossover = find_crossover_point(
b1=trend1[0], m1=trend1[1], b2=trend2[0], m2=trend2[1]
)
days_til_crossover = (
pd.to_datetime(crossover, unit="s") - pd.Timestamp.today()
).days
# add annotation with number of models and days til crossover
fig.add_annotation(
xref="paper",
yref="paper", # use paper coordinates
x=-0.05,
y=1.13,
text=f"Number of models: {len(filtered_df)}<br>Days til crossover: {days_til_crossover}",
showarrow=False,
font=dict(size=14, color="white"),
bgcolor="rgba(0,0,0,0.5)",
)
if toggle_annotations:
# get the points to annotate (only the highest rated model per month per license)
idx_to_annotate = filtered_df.groupby(["Month-Year", "License"])[
"rating"
].idxmax()
points_to_annotate_df = filtered_df.loc[idx_to_annotate]
for i, row in points_to_annotate_df.iterrows():
fig.add_annotation(
x=row["Release Date"],
y=row["rating"],
text=row["Model"],
showarrow=True,
arrowhead=0,
)
return fig, clean_df_for_display(filtered_df)
set_dark_mode = """
function refresh() {
const url = new URL(window.location);
if (url.searchParams.get('__theme') !== 'dark') {
url.searchParams.set('__theme', 'dark');
window.location.href = url.href;
}
}
"""
with gr.Blocks(
theme=gr.themes.Soft(
primary_hue=gr.themes.colors.sky,
secondary_hue=gr.themes.colors.green,
# spacing_size=gr.themes.sizes.spacing_sm,
text_size=gr.themes.sizes.text_sm,
font=[
gr.themes.GoogleFont("Open Sans"),
"ui-sans-serif",
"system-ui",
"sans-serif",
],
),
js=set_dark_mode,
) as demo:
gr.Markdown(
"""
<div style="text-align: center; max-width: 650px; margin: auto;">
<h1 style="font-weight: 900; margin-top: 5px;">🔬 Progress Tracker: Open vs. Proprietary LLMs 🔬</h1>
<p style="text-align: left; margin-top: 30px; margin-bottom: 30px; line-height: 20px;">
This app visualizes the progress of proprietary and open-source LLMs over time as scored by the <a href="https://leaderboard.lmsys.org/">LMSYS Chatbot Arena</a>.
The idea is inspired by <a href="https://www.linkedin.com/posts/maxime-labonne_arena-elo-graph-updated-with-new-models-activity-7187062633735368705-u2jB">this great work</a>
from <a href="https://huggingface.co/mlabonne/">Maxime Labonne</a>, and is intended to stay up-to-date as new models are released and evaluated.
<div style="text-align: left;">
<strong>Plot info:</strong>
<br>
<ul style="padding-left: 20px;">
<li> The ELO score (y-axis) is a measure of the relative strength of a model based on its performance against other models in the arena. </li>
<li> The Release Date (x-axis) corresponds to when the model was first publicly released or when its ELO results were first reported (for ease of automated updates). </li>
<li> Trend lines are based on Ordinary Least Squares (OLS) regression and adjust based on the filter criteria. </li>
<ul>
</div>
</p>
</div>
"""
)
with gr.Group():
with gr.Row(variant="compact"):
set_selector = gr.Dropdown(
choices=list(CAT_NAME_TO_EXPLANATION.keys()),
label="Select Category",
value="Overall",
info="Select the category to visualize",
)
min_score = gr.Slider(
minimum=min_elo_score,
maximum=max_elo_score,
value=(max_elo_score - min_elo_score) * 0.3 + min_elo_score,
step=50,
label="Minimum ELO Score",
info="Filter out low scoring models",
)
max_models_per_month = gr.Slider(
value=upper_models_per_month - 2,
minimum=1,
maximum=upper_models_per_month,
step=1,
label="Max Models per Month (per License)",
info="Limit to N best models per month per license to reduce clutter",
)
toggle_annotations = gr.Radio(
choices=[True, False],
label="Overlay Best Model Name",
value=True,
info="Toggle to overlay the name of the best model per month per license",
)
with gr.Row(variant="compact"):
with gr.Accordion("More options", open=False):
org_selector = gr.Dropdown(
choices=sorted(orgs),
label="Filter by Organization",
value=sorted(orgs),
multiselect=True,
info="Limit organizations included in plot",
)
# Show plot
filtered_df = gr.State()
with gr.Group():
with gr.Tab("Plot"):
plot = gr.Plot(show_label=False)
with gr.Tab("Raw Data"):
display_df = gr.DataFrame()
demo.load(
fn=filter_df,
inputs=[min_score, max_models_per_month, set_selector, org_selector],
outputs=filtered_df,
).then(
fn=build_plot,
inputs=[toggle_annotations, filtered_df],
outputs=[plot, display_df],
)
min_score.change(
fn=filter_df,
inputs=[min_score, max_models_per_month, set_selector, org_selector],
outputs=filtered_df,
).then(
fn=build_plot,
inputs=[toggle_annotations, filtered_df],
outputs=[plot, display_df],
)
max_models_per_month.change(
fn=filter_df,
inputs=[min_score, max_models_per_month, set_selector, org_selector],
outputs=filtered_df,
).then(
fn=build_plot,
inputs=[toggle_annotations, filtered_df],
outputs=[plot, display_df],
)
toggle_annotations.change(
fn=filter_df,
inputs=[min_score, max_models_per_month, set_selector, org_selector],
outputs=filtered_df,
).then(
fn=build_plot,
inputs=[toggle_annotations, filtered_df],
outputs=[plot, display_df],
)
set_selector.change(
fn=filter_df,
inputs=[min_score, max_models_per_month, set_selector, org_selector],
outputs=filtered_df,
).then(
fn=build_plot,
inputs=[toggle_annotations, filtered_df],
outputs=[plot, display_df],
)
org_selector.change(
fn=filter_df,
inputs=[min_score, max_models_per_month, set_selector, org_selector],
outputs=filtered_df,
).then(
fn=build_plot,
inputs=[toggle_annotations, filtered_df],
outputs=[plot, display_df],
)
gr.Markdown(
"""
<div style="text-align: center; max-width: 650px; margin: auto;">
<p style="margin-top: 40px;"> If you have any questions, feel free to open a discussion or <a href="https://twitter.com/andrewrreed">reach out to me on social</a>. </p>
</p>
</div>
"""
)
demo.launch()