CZ-EVAL / app.py
Hynek Kydlicek
new models
acc5a5f
raw
history blame
9.7 kB
"""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
import plotly.graph_objects as go
import pandas as pd
MODEL_NAME_COST = {
"anthropic/claude-2.1": 8,
"anthropic/claude-3-haiku": 0.25,
"anthropic/claude-3-opus": 15,
"anthropic/claude-3-sonnet": 3,
"cohere/command-r": 0.5,
"google/gemini-pro": 0.12,
"google/gemma-7b-it": 0.1,
"mistralai/mistral-large": 8,
"mistralai/mistral-medium": 2.7,
"mistralai/mixtral-8x7b-instruct": 0.7,
"openai/gpt-3.5-turbo": 0.5,
"openai/gpt-4-1106-preview": 10,
}
def make_default_md():
leaderboard_md = f"""
# ๐Ÿ† CZ-EVAL Leaderboard
[Developer](https://me.hynky.name/) | [Twitter](https://twitter.com/HKydlicek)
CZ-EVAL is a evaluation leadboard of Tasks in Czech for LLMs.
It's evaluated on following datasets:
- Math Problems Understanding [Klokan-QA](https://huggingface.co/datasets/hynky/klokan-qa)
- Reasoning and General Knowledge [TSP-QA](https://huggingface.co/datasets/hynky/tsp-qa)
๐Ÿ’ป Code: The evaluation code can be found at [hynky1999/LLM-Eval](https://github.com/hynky1999/LLM-Eval). Model inference is done using [Open-Router](https://openrouter.ai/) or on cloud using [Modal Labs](https://modal.com/).
"""
return leaderboard_md
def make_arena_leaderboard_md(arena_df):
total_models = len(arena_df)
leaderboard_md = f"""
Total #models: **{total_models}**. Last updated: Mar 17, 2024.
"""
return leaderboard_md
def make_full_leaderboard_md(elo_results):
leaderboard_md = f"""
Three benchmarks are displayed: **Arena Elo**, **MT-Bench** and **MMLU**.
- [Klokan-QA](https://huggingface.co/datasets/hynky/klokan-qa) - Mathematical competitions dataset
- [TSP](https://huggingface.co/datasets/hynky/TSP) - Comprehensive dataset of
"""
return leaderboard_md
# Combine all category accuracies into a single DataFrame
def plot_spider(df, title):
categories = df.columns.tolist()[1:]
categories = [
*categories,
categories[0],
] # Ensure the graph is circular by appending the start to the end
colors = [
'#1f77b4', # muted blue
'#ff7f0e', # safety orange
'#2ca02c', # cooked asparagus green
'#d62728', # brick red
'#9467bd', # muted purple
'#8c564b', # chestnut brown
'#e377c2', # raspberry yogurt pink
'#7f7f7f', # middle gray
'#bcbd22', # curry yellow-green
'#17becf', # blue-teal
'#f7b6d2', # pastel pink
'#bcbd22', # faded green
'#dbdb8d', # light olive
'#17becf', # soft blue
'#9edae5', # light blue
'#c5b0d5', # soft purple
'#c49c94', # dusty rose
'#f7b6d2', # pastel pink
'#bcbd22', # faded green
'#dbdb8d', # light olive
'#17becf', # soft blue
'#9edae5', # light blue
'#c5b0d5', # soft purple
'#c49c94', # dusty rose
]
# Setting for 1000x1000
fig_1000 = go.Figure()
for i, (idx, row) in enumerate(df.iterrows()):
name = row[0]
row = row.tolist()[1:]
row = row + [
row[0]
] # Ensure the graph is circular by appending the start to the end
color = colors[i]
fig_1000.add_trace(
go.Scatterpolar(
r=row,
theta=categories,
opacity=0.4,
name=name,
line=dict(
color=color, width=4
), # Adjust line width for better visibility
)
)
fig_1000.update_layout(
width=600,
height=950,
polar=dict(
angularaxis=dict(
gridwidth=2, # Increase line width for better visibility
rotation=90,
direction="clockwise",
),
radialaxis=dict(
visible=True,
range=[0, 100],
angle=45,
tickangle=45,
tickvals=[0, 25, 50, 75, 100],
ticktext=["0%", "25%", "50%", "75%", "100%"],
),
),
title_text=title,
title_x=0.5,
title_y=0.97,
title_xanchor="center",
title_yanchor="top",
title_font_size=24,
title_font_color="#333333",
font=dict(family="Arial", size=16, color="#333333"),
legend=dict(
orientation="h", yanchor="bottom", y=-0.45, xanchor="center", x=0.5
),
)
return fig_1000
def openrouter_hyperlink(model_name):
return f'<a target="_blank" href="https://openrouter.ai/models/{model_name}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
def get_full_table(model_table_df):
num_cols = ["klokan", "culture", "analytical", "critical", "verbal"]
# Multiply by 100 and round to 2 decimals
# Add average
model_table_df["average"] = model_table_df[num_cols].mean(axis=1)
model_table_df[num_cols + ["average"]] = model_table_df[
num_cols + ["average"]
].apply(lambda x: round(x * 100, 2))
# Sort and add rank
model_table_df.sort_values(by="average", ascending=False, inplace=True)
model_table_df.insert(0, "rank", np.arange(1, len(model_table_df) + 1))
# Add cost
model_table_df["completion_price"] = model_table_df["model_name"].apply(
lambda x: f"{MODEL_NAME_COST.get(x, "N/A")}$"
)
# Add link
model_table_df["model_name"] = model_table_df["model_name"].apply(
lambda x: openrouter_hyperlink(x)
)
# Ensure the dataframe is in the correct order before renaming
model_table_df = model_table_df[["rank", "model_name", "completion_price", "klokan", "culture", "analytical", "critical", "verbal", "average"]]
model_table_df.rename(
columns={
"model_name": "๐Ÿค– Model",
"completion_price": "๐Ÿ’ฐ Cost (1M-Tokens)",
"klokan": "๐Ÿงฎ Klokan-QA",
"culture": "๐ŸŒ TSP-Culture",
"analytical": "๐Ÿ” TSP-Analytical",
"critical": "๐Ÿ’ก TSP-Critical",
"verbal": "๐Ÿ“– TSP-Verbal",
"average": "๐Ÿ“Š Average",
},
inplace=True,
)
return model_table_df
def build_leaderboard_tab(leaderboard_table_file, klokan_table_file, tsp_table_file):
results = pd.read_csv(leaderboard_table_file)
results = get_full_table(results)
# p1, p2 = get_grafs(pd.read_json(klokan_table_file), pd.read_json(tsp_table_file))
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("CZ-EVAL Leaderboard", id=0):
md = make_arena_leaderboard_md(results)
gr.Markdown(md, elem_id="leaderboard_markdown")
gr.Dataframe(
datatype=[
"str",
"markdown",
"number",
"number",
"number",
"number",
"number",
"number",
"str",
"str",
"str",
],
value=results,
elem_id="arena_leaderboard_dataframe",
height=700,
column_widths=[
70,
200,
110,
120,
120,
120,
120,
100,
100,
],
wrap=True,
)
p1 = plot_spider(pd.read_csv(klokan_table_file), "Klokan-QA - Acurracy")
p2 = plot_spider(pd.read_csv(tsp_table_file), "TSP - Accuracy")
gr.Markdown(
f"""## More Statistics for CZ-EVAL\n
Below are figures for more statistics.
""",
elem_id="leaderboard_markdown",
)
with gr.Row():
with gr.Column():
gr.Markdown(
"#### Figure 1: Performance of models on Klokan-QA per difficulty"
)
plot_1 = gr.Plot(p1, show_label=False)
with gr.Column():
gr.Markdown("#### Figure 2: Performance of models on TSP dataset")
plot_2 = gr.Plot(p2, show_label=False)
return [md_1, plot_1, plot_2]
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%;
}
"""
def build_demo(leadboard_table, klokan_table, tsp_table):
text_size = gr.themes.sizes.text_lg
with gr.Blocks(
title="CZ-EVAL Leaderboard",
theme=gr.themes.Base(text_size=text_size),
css=block_css,
) as demo:
leader_components = build_leaderboard_tab(
leadboard_table, klokan_table, tsp_table
)
return demo
demo = build_demo(
leadboard_table="./leaderboard/table.csv",
klokan_table="./leaderboard/klokan.csv",
tsp_table="./leaderboard/tsp.csv",
)
if __name__ == "__main__":
demo.launch()