clem-leaderboard / src /trend_utils.py
sherzod-hakimov's picture
trend plot
54f2589
raw
history blame
18 kB
## Fetch Model Registry and clemscores
import requests
import pandas as pd
from datetime import datetime
import pandas as pd
import plotly.express as px
import plotly.graph_objects as go
import numpy as np
from src.assets.text_content import REGISTRY_URL, REPO, BENCHMARK_FILE
from src.leaderboard_utils import get_github_data
# Cut-off date from where to start the trendgraph
START_DATE = '2023-06-01'
def get_param_size(params: str) -> float:
"""Convert parameter size from string to float.
Args:
params (str): The parameter size as a string (e.g., '1000B', '1T').
Returns:
float: The size of parameters in float.
"""
if not params:
param_size = 0
else:
if params[-1] == "B":
param_size = params[:-1]
param_size = float(param_size)
elif params[-1] == "T":
param_size = params[:-1]
param_size = float(param_size)
param_size *= 1000
else:
print("Not a valid parameter size")
return param_size
def date_difference(date_str1: str, date_str2: str) -> int:
"""Calculate the difference in days between two dates.
Args:
date_str1 (str): The first date as a string in 'YYYY-MM-DD' format.
date_str2 (str): The second date as a string in 'YYYY-MM-DD' format.
Returns:
int: The difference in days between the two dates.
"""
date_format = "%Y-%m-%d"
date1 = datetime.strptime(date_str1, date_format)
date2 = datetime.strptime(date_str2, date_format)
return (date1 - date2).days
def populate_list(df: pd.DataFrame, abs_diff: float) -> list:
"""Create a list of models based on clemscore differences.
Args:
df (pd.DataFrame): DataFrame containing model data.
abs_diff (float): The absolute difference threshold for clemscore.
Returns:
list: A list of model names that meet the criteria.
"""
l = [df.iloc[0]['model']]
prev_clemscore = df.iloc[0]['clemscore']
prev_date = df.iloc[0]['release_date']
for i in range(1, len(df)):
curr_clemscore = df.iloc[i]['clemscore']
curr_date = df.iloc[i]['release_date']
date_diff = date_difference(curr_date, prev_date)
if curr_clemscore - prev_clemscore >= abs_diff:
if date_diff == 0:
l[-1] = df.iloc[i]['model']
else:
l.append(df.iloc[i]['model'])
prev_clemscore = curr_clemscore
prev_date = curr_date
# # Add the last model if the difference between the last and previous date is greater than 15 days
# last_date = df.iloc[-1]['release_date']
# if date_difference(last_date, prev_date) > 15:
# l.append(df.iloc[-1]['model'])
return l
def get_models_to_display(result_df: pd.DataFrame, open_dip: float = 0, comm_dip: float = 0) -> tuple:
"""Retrieve models to display based on clemscore differences.
Args:
result_df (pd.DataFrame): DataFrame containing model data.
open_dip (float, optional): Threshold for open models. Defaults to 0.
comm_dip (float, optional): Threshold for commercial models. Defaults to 0.
Returns:
tuple: Two lists of model names (open and commercial).
"""
open_model_df = result_df[result_df['open_weight']==True]
comm_model_df = result_df[result_df['open_weight']==False]
open_model_df = open_model_df.sort_values(by='release_date', ascending=True)
comm_model_df = comm_model_df.sort_values(by='release_date', ascending=True)
open_models = populate_list(open_model_df, open_dip)
comm_models = populate_list(comm_model_df, comm_dip)
return open_models, comm_models
def get_trend_data(text_dfs: list, model_registry_data: list) -> pd.DataFrame:
"""Process text data frames to extract model information.
Args:
text_dfs (list): List of DataFrames containing model information.
model_registry_data (list): List of dictionaries containing model registry data.
Returns:
pd.DataFrame: DataFrame containing processed model data.
"""
visited = set() # Track models that have been processed
result_df = pd.DataFrame(columns=['model', 'clemscore', 'open_weight', 'release_date', 'parameters', 'est_flag'])
for df in text_dfs:
for i in range(len(df)):
model_name = df['Model'].iloc[i]
if model_name not in visited:
visited.add(model_name)
for dict_obj in model_registry_data:
if dict_obj["model_name"] == model_name:
if dict_obj["parameters"] == "" :
params = "1000B"
est_flag = True
else:
params = dict_obj['parameters']
est_flag = False
param_size = get_param_size(params)
new_data = {'model': model_name, 'clemscore': df['Clemscore'].iloc[i], 'open_weight':dict_obj['open_weight'],
'release_date': dict_obj['release_date'], 'parameters': param_size, 'est_flag': est_flag}
result_df.loc[len(result_df)] = new_data
break
return result_df # Return the compiled DataFrame
def get_plot(df: pd.DataFrame, start_date: str = '2023-06-01', end_date: str = '2024-12-30',
benchmark_ticks: dict = {}, benchmark_update = {}, **plot_kwargs) -> go.Figure:
"""Generate a scatter plot for the given DataFrame.
Args:
df (pd.DataFrame): DataFrame containing model data.
start_date (str, optional): Start date for filtering. Defaults to '2023-06-01'.
end_date (str, optional): End date for filtering. Defaults to '2024-12-30'.
benchmark_ticks (dict, optional): Custom benchmark ticks for the version dates. Defaults to {}.
benchmark_update (dict, optional): Custom benchmark metadata containing last_updated date for the versions. Defaults to {}.
Keyword Args:
open_dip (float, optional): Threshold for open models' clemscore differences. Max dip in clemscore allowed to be considered in trend.
comm_dip (float, optional): Threshold for commercial models' clemscore differences. Max dip in clemscore allowed to be considered in trend.
height (int, optional): Height of the plot in pixels. Adjusted for mobile or desktop views.
mobile_view (bool, optional): Flag to indicate if the plot should be optimized for mobile display. Defaults to False.
Returns:
go.Figure: The generated plot.
"""
open_dip = plot_kwargs['open_dip']
comm_dip = plot_kwargs['comm_dip']
height = plot_kwargs['height']
width = plot_kwargs['width']
mobile_view = True if plot_kwargs['mobile_view'] else False
max_clemscore = df['clemscore'].max()
# Convert 'release_date' to datetime
df['Release date'] = pd.to_datetime(df['release_date'], format='ISO8601')
# Filter out data before April 2023/START_DATE
df = df[df['Release date'] >= pd.to_datetime(start_date)]
open_model_list, comm_model_list = get_models_to_display(df, open_dip, comm_dip)
models_to_display = open_model_list + comm_model_list
print(f"open_model_list: {open_model_list}, comm_model_list: {comm_model_list}")
# Create a column to indicate if the model should be labeled
df['label_model'] = df['model'].apply(lambda x: x if x in models_to_display else "")
# If mobile_view, then show only the models in models_to_display i.e. on the trend line #minimalistic
if mobile_view:
df = df[df['model'].isin(models_to_display)]
# Add an identifier column to each DataFrame
df['Model Type'] = df['open_weight'].map({True: 'Open-Weight', False: 'Commercial'})
marker_size = df['parameters'].apply(lambda x: np.sqrt(x) if x > 0 else np.sqrt(400)).astype(float) # Arbitrary sqrt value to scale marker size based on parameter size
open_color = 'red'
comm_color = 'blue'
# Create the scatter plot
fig = px.scatter(df,
x="Release date",
y="clemscore",
color="Model Type", # Differentiates the datasets by color
hover_name="model",
size=marker_size,
size_max=40, # Max size of the circles
template="plotly_white",
hover_data={ # Customize hover information
"Release date": True, # Show the release date
"clemscore": True, # Show the clemscore
"Model Type": True # Show the model type
},
custom_data=["model", "Release date", "clemscore"] # Specify custom data columns for hover
)
fig.update_traces(
hovertemplate='Model Name: %{customdata[0]}<br>Release date: %{customdata[1]}<br>Clemscore: %{customdata[2]}<br>'
)
# Sort dataframes for line plotting
df_open = df[df['model'].isin(open_model_list)].sort_values(by='Release date')
df_commercial = df[df['model'].isin(comm_model_list)].sort_values(by='Release date')
## Custom tics for x axis
# Define the start and end dates
start_date = pd.to_datetime(start_date)
end_date = pd.to_datetime(end_date)
# Generate ticks every two months
date_range = pd.date_range(start=start_date, end=end_date, freq='2MS') # '2MS' stands for 2 Months Start frequency
# Create labels for these ticks
custom_ticks = {date: date.strftime('%b %Y') for date in date_range}
## Benchmark Version ticks
benchmark_tickvals = list(pd.to_datetime(list(benchmark_ticks.keys())))
custom_ticks = {k:v for k,v in custom_ticks.items() if k not in benchmark_tickvals}
custom_tickvals = list(custom_ticks.keys())
for date, version in benchmark_ticks.items():
# Find the corresponding update date from benchmark_update based on the version name
update_date = next((update_date for update_date, ver in benchmark_update.items() if version in ver), None)
if update_date:
# Add vertical black dotted line for each benchmark_tick date
fig.add_shape(
go.layout.Shape(
type='line',
x0=date,
x1=date,
y0=0,
y1=1,
yref='paper',
line=dict(color='#A9A9A9', dash='dash'), # Black dotted line
)
)
# Add hover information across the full y-axis range
fig.add_trace(
go.Scatter(
x=[date]*100,
y=list(range(0,100)), # Covers full y-axis range
mode='markers',
line=dict(color='rgba(255,255,255,0)', width=0), # Fully transparent line
hovertext=[
f"Version: {version} released on {date.strftime('%d %b %Y')}, last updated on: {update_date.strftime('%d %b %Y')}"
for _ in range(100)
], # Unique hovertext for all points
hoverinfo="text",
hoveron='points',
showlegend=False
)
)
if mobile_view:
# Remove custom_tickvals within -1 month to +1 month of benchmark_tickvals for better visibility
one_month = pd.DateOffset(months=1)
filtered_custom_tickvals = [
date for date in custom_tickvals
if not any((benchmark_date - one_month <= date <= benchmark_date + one_month) for benchmark_date in benchmark_tickvals)
]
# Alternate <br> for benchmark ticks based on date difference (Eg. v1.6, v1.6.5 too close to each other for MM benchmark)
benchmark_tick_texts = []
for i in range(len(benchmark_tickvals)):
if i == 0:
benchmark_tick_texts.append(f"<br><br><b>{benchmark_ticks[benchmark_tickvals[i]]}</b>")
else:
date_diff = (benchmark_tickvals[i] - benchmark_tickvals[i - 1]).days
if date_diff <= 75:
benchmark_tick_texts.append(f"<br><br><br><b>{benchmark_ticks[benchmark_tickvals[i]]}</b>")
else:
benchmark_tick_texts.append(f"<br><br><b>{benchmark_ticks[benchmark_tickvals[i]]}</b>")
fig.update_xaxes(
tickvals=filtered_custom_tickvals + benchmark_tickvals, # Use filtered_custom_tickvals
ticktext=[f"{date.strftime('%b')}<br>{date.strftime('%y')}" for date in filtered_custom_tickvals] +
benchmark_tick_texts, # Use the new benchmark tick texts
tickangle=0,
tickfont=dict(size=10)
)
fig.update_yaxes(range=[0, 110]) # Set y-axis range to 110 for better visibility of legend and avoiding overlap with interactivity block of plotly on top-right
display_mode = 'lines+markers'
else:
fig.update_xaxes(
tickvals=custom_tickvals + benchmark_tickvals, # Use filtered_custom_tickvals
ticktext=[f"{date.strftime('%b')} {date.strftime('%Y')}" for date in custom_tickvals] +
[f"<br><span style='font-size:12px;'><b>{benchmark_ticks[date]}</b></span>" for date in benchmark_tickvals], # Added <br> for vertical alignment
tickangle=0,
tickfont=dict(size=10)
)
fig.update_yaxes(range=[0, max_clemscore+10])
display_mode = 'lines+markers+text'
# Add lines connecting the points for open models
fig.add_trace(go.Scatter(x=df_open['Release date'], y=df_open['clemscore'],
mode=display_mode, # Include 'text' in the mode
name='Open Models Trendline',
text=df_open['label_model'], # Use label_model for text labels
textposition='top center', # Position of the text labels
line=dict(color=open_color), showlegend=False))
# Add lines connecting the points for commercial models
fig.add_trace(go.Scatter(x=df_commercial['Release date'], y=df_commercial['clemscore'],
mode=display_mode, # Include 'text' in the mode
name='Commercial Models Trendline',
text=df_commercial['label_model'], # Use label_model for text labels
textposition='top center', # Position of the text labels
line=dict(color=comm_color), showlegend=False))
# Update layout to ensure text labels are visible
fig.update_traces(textposition='top center')
# Update the Legend Position and plot dimensions
fig.update_layout(height=height,
legend=dict(
yanchor="top",
y=0.99,
xanchor="left",
x=0.01
)
)
if width:
print("Custom Seting the Width :")
fig.update_layout(width=width)
return fig
def get_final_trend_plot(benchmark: str = "Text", mobile_view: bool = False) -> go.Figure:
"""Fetch and generate the final trend plot for all models.
Args:
benchmark (str, optional): The benchmark type to use. Defaults to "Text".
mobile_view (bool, optional): Flag to indicate mobile view. Defaults to False.
Returns:
go.Figure: The generated trend plot for selected benchmark.
"""
# Fetch Model Registry
response = requests.get(REGISTRY_URL)
model_registry_data = response.json()
# Custom tick labels
json_url = REPO + BENCHMARK_FILE
response = requests.get(json_url)
# Check if the JSON file request was successful
if response.status_code != 200:
print(f"Failed to read JSON file: Status Code: {response.status_code}")
json_data = response.json()
versions = json_data['versions']
if mobile_view:
height = 450
width = 375
else:
height = 1000
width = None
plot_kwargs = {'height': height, 'width': width, 'open_dip': 0, 'comm_dip': 0,
'mobile_view': mobile_view}
benchmark_ticks = {}
benchmark_update = {}
if benchmark == "Text":
text_dfs = get_github_data()['text']['dataframes']
text_result_df = get_trend_data(text_dfs, model_registry_data)
## Get benchmark tickvalues as dates for X-axis
for ver in versions:
if 'multimodal' not in ver['version']: # Skip MM specific benchmark dates
benchmark_ticks[pd.to_datetime(ver['release_date'])] = ver['version']
if pd.to_datetime(ver['last_updated']) not in benchmark_update:
benchmark_update[pd.to_datetime(ver['last_updated'])] = [ver['version']]
else:
benchmark_update[pd.to_datetime(ver['last_updated'])].append(ver['version'])
fig = get_plot(text_result_df, start_date=START_DATE, end_date=datetime.now().strftime('%Y-%m-%d'), benchmark_ticks=benchmark_ticks, benchmark_update=benchmark_update, **plot_kwargs)
else:
mm_dfs = get_github_data()['multimodal']['dataframes']
result_df = get_trend_data(mm_dfs, model_registry_data)
df = result_df
for ver in versions:
if 'multimodal' in ver['version']:
temp_ver = ver['version']
temp_ver = temp_ver.replace('_multimodal', '')
benchmark_ticks[pd.to_datetime(ver['release_date'])] = temp_ver ## MM benchmark dates considered after v1.6 (incl.)
benchmark_update[pd.to_datetime(ver['last_updated'])] = temp_ver
fig = get_plot(df, start_date=START_DATE, end_date=datetime.now().strftime('%Y-%m-%d'), benchmark_ticks=benchmark_ticks, benchmark_update=benchmark_update, **plot_kwargs)
return fig