clem-leaderboard / src /plot_utils.py
sherzod-hakimov's picture
multimodal leaderboard
923aff9
raw
history blame contribute delete
No virus
10.7 kB
import pandas as pd
import plotly.express as px
import requests
import json
import gradio as gr
from src.assets.text_content import SHORT_NAMES, TEXT_NAME, MULTIMODAL_NAME
from src.leaderboard_utils import get_github_data
def plotly_plot(df: pd.DataFrame, list_op: list, list_co: list,
show_all: list, show_names: list, show_legend: list,
mobile_view: list):
"""
Takes in a list of models for a plotly plot
Args:
df: A dummy dataframe of latest version
list_op: The list of open source models to show in the plot, updated from frontend
list_co: The list of commercial models to show in the plot, updated from frontend
show_all: Either [] or ["Show All Models"] - toggle view to plot all models
show_names: Either [] or ["Show Names"] - toggle view to show model names on plot
show_legend: Either [] or ["Show Legend"] - toggle view to show legend on plot
mobile_view: Either [] or ["Mobile View"] - toggle view to for smaller screens
Returns:
Fig: plotly figure of % played v/s quality score
"""
LIST = list_op + list_co
# Get list of all models and append short names column to df
list_columns = list(df.columns)
ALL_LIST = list(df[list_columns[0]].unique())
short_names = label_map(ALL_LIST)
list_short_names = list(short_names.values())
df["Short"] = list_short_names
if show_all:
LIST = ALL_LIST
# Filter dataframe based on the provided list of models
df = df[df[list_columns[0]].isin(LIST)]
if show_names:
fig = px.scatter(df, x=list_columns[2], y=list_columns[3], color=list_columns[0], symbol=list_columns[0],
color_discrete_map={"category1": "blue", "category2": "red"},
hover_name=list_columns[0], template="plotly_white", text="Short")
fig.update_traces(textposition='top center')
else:
fig = px.scatter(df, x=list_columns[2], y=list_columns[3], color=list_columns[0], symbol=list_columns[0],
color_discrete_map={"category1": "blue", "category2": "red"},
hover_name=list_columns[0], template="plotly_white")
if not show_legend:
fig.update_layout(showlegend=False)
fig.update_layout(
xaxis_title='% Played',
yaxis_title='Quality Score',
title='Overview of benchmark results',
height=1000
)
fig.update_xaxes(range=[-5, 105])
fig.update_yaxes(range=[-5, 105])
if mobile_view:
fig.update_layout(height=300)
if mobile_view and show_legend:
fig.update_layout(height=450)
fig.update_layout(legend=dict(
yanchor="bottom",
y=-5.52,
xanchor="left",
x=0.01
))
fig.update_layout(
xaxis_title="",
yaxis_title="",
title="% Played v/s Quality Score"
)
return fig
def shorten_model_name(full_name):
# Split the name into parts
parts = full_name.split('-')
# Process the name parts to keep only the parts with digits (model sizes and versions)
short_name_parts = [part for part in parts if any(char.isdigit() for char in part)]
if len(parts) == 1:
short_name = ''.join(full_name[0:min(3, len(full_name))])
else:
# Join the parts to form the short name
short_name = '-'.join(short_name_parts)
# Remove any leading or trailing hyphens
short_name = full_name[0] + '-' + short_name.strip('-')
return short_name
def label_map(model_list: list) -> dict:
"""
Generate a map from long names to short names, to plot them in frontend graph
Define the short names in src/assets/text_content.py
Args:
model_list: A list of long model names
Returns:
short_name: A dict from long to short name
"""
short_names = {}
for model_name in model_list:
if model_name in SHORT_NAMES:
short_name = SHORT_NAMES[model_name]
else:
short_name = shorten_model_name(model_name)
# Define the short name and indicate both models are same
short_names[model_name] = short_name
return short_names
def split_models(model_list: list):
"""
Split the models into open source and commercial
"""
open_models = []
commercial_models = []
open_backends = {"huggingface_local", "huggingface_multimodal", "openai_compatible"} # Define backends considered as open
# Load model registry data from main repo
model_registry_url = "https://raw.githubusercontent.com/clp-research/clembench/main/backends/model_registry.json"
response = requests.get(model_registry_url)
if response.status_code == 200:
json_data = json.loads(response.text)
# Classify as Open or Commercial based on the defined backend in the model registry
backend_mapping = {}
for model_name in model_list:
model_prefix = model_name.split('-')[0] # Get the prefix part of the model name
for entry in json_data:
if entry["model_name"].startswith(model_prefix):
backend = entry["backend"]
# Classify based on backend
if backend in open_backends:
open_models.append(model_name)
else:
commercial_models.append(model_name)
break
else:
print(f"Failed to read JSON file: Status Code : {response.status_code}")
open_models.sort(key=lambda o: o.upper())
commercial_models.sort(key=lambda c: c.upper())
# Add missing model from the model_registry
if "dolphin-2.5-mixtral-8x7b" in model_list:
open_models.append("dolphin-2.5-mixtral-8x7b")
return open_models, commercial_models
"""
Update Functions, for when the leaderboard selection changes
"""
def update_open_models(leaderboard: str = TEXT_NAME):
"""
Change the checkbox group of Open Models based on the leaderboard selected
Args:
leaderboard: Selected leaderboard from the frontend [Default - Text Leaderboard]
Return:
Updated checkbox group for Open Models, based on the leaderboard selected
"""
github_data = get_github_data()
leaderboard_data = github_data["text" if leaderboard == TEXT_NAME else "multimodal"][0]
models = leaderboard_data.iloc[:, 0].unique().tolist()
open_models, commercial_models = split_models(models)
return gr.CheckboxGroup(
open_models,
value=[],
elem_id="value-select-1",
interactive=True,
)
def update_closed_models(leaderboard: str = TEXT_NAME):
"""
Change the checkbox group of Closed Models based on the leaderboard selected
Args:
leaderboard: Selected leaderboard from the frontend [Default - Text Leaderboard]
Return:
Updated checkbox group for Closed Models, based on the leaderboard selected
"""
github_data = get_github_data()
leaderboard_data = github_data["text" if leaderboard == TEXT_NAME else "multimodal"][0]
models = leaderboard_data.iloc[:, 0].unique().tolist()
open_models, commercial_models = split_models(models)
return gr.CheckboxGroup(
commercial_models,
value=[],
elem_id="value-select-2",
interactive=True,
)
def get_plot_df(leaderboard: str = TEXT_NAME) -> pd.DataFrame:
"""
Get the DataFrame for plotting based on the selected leaderboard.
Args:
leaderboard: Selected leaderboard.
Returns:
DataFrame with model data.
"""
github_data = get_github_data()
return github_data["text" if leaderboard == TEXT_NAME else "multimodal"][0]
"""
Reset Functions for when the Leaderboard selection changes
"""
def reset_show_all():
return gr.CheckboxGroup(
["Select All Models"],
label="Show plot for all models πŸ€–",
value=[],
elem_id="value-select-3",
interactive=True,
)
def reset_show_names():
return gr.CheckboxGroup(
["Show Names"],
label="Show names of models on the plot 🏷️",
value=[],
elem_id="value-select-4",
interactive=True,
)
def reset_show_legend():
return gr.CheckboxGroup(
["Show Legend"],
label="Show legend on the plot πŸ’‘",
value=[],
elem_id="value-select-5",
interactive=True,
)
def reset_mobile_view():
return gr.CheckboxGroup(
["Mobile View"],
label="View plot on smaller screens πŸ“±",
value=[],
elem_id="value-select-6",
interactive=True,
)
if __name__ == '__main__':
mm_model_list = ['gpt-4o-2024-05-13', 'gpt-4-1106-vision-preview', 'claude-3-opus-20240229', 'gemini-1.5-pro-latest',
'gemini-1.5-flash-latest', 'llava-v1.6-34b-hf', 'llava-v1.6-vicuna-13b-hf', 'idefics-80b-instruct',
'llava-1.5-13b-hf', 'idefics-9b-instruct']
text_model_list = ['vicuna-33b-v1.3', 'gpt-4-0125-preview', 'gpt-4-turbo-2024-04-09', 'claude-3-5-sonnet-20240620', 'gpt-4-1106-preview',
'gpt-4-0613', 'gpt-4o-2024-05-13', 'claude-3-opus-20240229', 'gemini-1.5-pro-latest',
'Meta-Llama-3-70B-Instruct-hf', 'claude-2.1', 'gemini-1.5-flash-latest', 'claude-3-sonnet-20240229',
'Qwen1.5-72B-Chat', 'mistral-large-2402', 'gpt-3.5-turbo-0125', 'gemini-1.0-pro', 'command-r-plus', 'openchat_3.5',
'claude-3-haiku-20240307', 'sheep-duck-llama-2-70b-v1.1', 'Meta-Llama-3-8B-Instruct-hf', 'openchat-3.5-1210',
'WizardLM-70b-v1.0', 'openchat-3.5-0106', 'Qwen1.5-14B-Chat', 'mistral-medium-2312', 'Qwen1.5-32B-Chat',
'codegemma-7b-it', 'dolphin-2.5-mixtral-8x7b', 'CodeLlama-34b-Instruct-hf', 'command-r', 'gemma-1.1-7b-it',
'SUS-Chat-34B', 'Mixtral-8x22B-Instruct-v0.1', 'tulu-2-dpo-70b', 'Nous-Hermes-2-Mixtral-8x7B-SFT',
'WizardLM-13b-v1.2', 'Mistral-7B-Instruct-v0.2', 'Yi-34B-Chat', 'Mixtral-8x7B-Instruct-v0.1',
'Mistral-7B-Instruct-v0.1', 'Yi-1.5-34B-Chat', 'vicuna-13b-v1.5', 'Yi-1.5-6B-Chat', 'Starling-LM-7B-beta',
'sheep-duck-llama-2-13b', 'Yi-1.5-9B-Chat', 'gemma-1.1-2b-it', 'Qwen1.5-7B-Chat', 'gemma-7b-it',
'llama-2-70b-chat-hf', 'Qwen1.5-0.5B-Chat', 'Qwen1.5-1.8B-Chat']
om, cm = split_models(mm_model_list)
print("Open")
print(om)
print("Closed")
print(cm)