import gradio as gr
from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
import pandas as pd
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import snapshot_download
from src.about import CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE
from src.tasks import TASK_DESCRIPTIONS, MEASURE_DESCRIPTION
from src.display.css_html_js import custom_css
from src.display.utils import BENCHMARK_COLS, COLS, EVAL_COLS, EVAL_TYPES, AutoEvalColumn, ModelType, fields, WeightType, Precision
from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN
from src.populate import get_evaluation_queue_df, get_leaderboard_df
from src.submission.submit import add_new_eval
import random
import matplotlib.pyplot as plt
import re
import plotly.express as px
import plotly.graph_objects as go
import numpy as np
def mean_of_max_per_field(df):
"""
Calcola il massimo per ciascun campo e poi la media dei massimi.
Args:
df (pd.DataFrame): DataFrame con colonne TE, SA, HS, AT, WIC, FAQ, LS, SU, NER, REL
Returns:
float: media dei valori massimi dei campi
"""
fields = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
#print(df.columns)
# Controlla che tutte le colonne esistano nel DataFrame
missing = [f for f in fields if f not in df.columns]
if missing:
raise ValueError(f"Le seguenti colonne mancano nel DataFrame: {missing}")
# Calcola il massimo per ciascun campo
max_values = df[fields].max()
# Calcola la media dei massimi
mean_max = max_values.mean()
return mean_max
def barplot_mean_few_minus_zero_shot(dataframe, tasks=None):
if tasks is None:
tasks = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
task_means = {}
for task in tasks:
if task not in dataframe.columns:
continue
# Separa few-shot e zero-shot
few_shot = dataframe[dataframe['IS_FS'] == True][["Model", task]]
zero_shot = dataframe[dataframe['IS_FS'] == False][["Model", task]]
# Allinea i modelli
merged = pd.merge(few_shot, zero_shot, on="Model", suffixes=("_few", "_zero"))
# Rimuovi righe con valori mancanti
merged = merged.dropna(subset=[f"{task}_few", f"{task}_zero"])
if merged.empty:
continue
# Calcola differenza few - zero
diff = merged[f"{task}_few"] - merged[f"{task}_zero"]
# Calcola la media
task_means[task] = diff.mean()
# Crea barplot
fig = go.Figure([go.Bar(
x=list(task_means.keys()),
y=list(task_means.values()),
marker_color="#ff7f0e",
text=[f"{v:.2f}" for v in task_means.values()],
textposition="outside",
hovertemplate="%{x}
Mean Delta Accuracy: %{y:.2f}%
"
"Only in Summarization (SU) does it provide no accuracy gain.",
xref="paper", yref="paper",
x=0, y=-0.2,
showarrow=False,
font=dict(size=11, color="gray"),
align="left"
)
return fig
def boxplot_per_task(dataframe=None, baselines=None, references=None):
#print(dataframe.columns)
tasks = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
if dataframe is None:
np.random.seed(42)
dataframe = pd.DataFrame({
task: np.random.uniform(0.4, 0.9, 20) * 100
for task in tasks
})
if baselines is None:
baselines = {task: np.random.randint(50, 70) for task in tasks}
colors = ["#1f77b4", "#ff7f0e", "#2ca02c", "#d62728", "#9467bd",
"#8c564b", "#e377c2", "#7f7f7f", "#bcbd22", "#17becf"]
fig = go.Figure()
for i, task in enumerate(tasks):
if task in dataframe.columns:
y_data = dataframe[task].dropna().tolist()
# boxplot
fig.add_trace(go.Box(
y=y_data,
name=task,
marker=dict(color=colors[i]),
line=dict(color="black", width=2),
fillcolor=colors[i],
opacity=0.7,
hovertemplate=""+task+"
Accuracy: %{y:.2f}%
"
"models at EVALITA (dashed black line); in NER and REL they remain lower.
"
"Dashed red lines show GPT-4o reference results for generative tasks."
),
xref="paper", yref="paper",
x=0.5, y=-0.30,
showarrow=False,
font=dict(size=11, color="gray"),
align="left"
)
fig.update_yaxes(range=[0, 100], fixedrange=True)
return fig
# EVALITA results
BASELINES = {
"TE":71.00, "SA": 66.38, "HS": 80.88, "AT": 82.40, "WIC": 85.00,
"LS": 38.82, "SU": 38.91, "NER":88.00, "REL": 62.99
}
# GPT-4o
REFERENCES = {
"NER": 79.11,
"REL": 63.32,
"LS": 59.25,
"SU": 33.04
}
def boxplot_prompts_per_task(dataframe, tasks=None):
if tasks is None:
tasks = ["TE", "SA", "HS", "AT", "WIC", "FAQ", "LS", "SU", "NER", "REL"]
# Lista delle colonne da aggiornare
cols_to_update = ["REL Best Prompt Id", "NER Best Prompt Id", "SU Best Prompt Id", "LS Best Prompt Id"]
# Applichiamo la trasformazione
for col in cols_to_update:
dataframe[col] = dataframe[col].replace({1: 7, 2: 8})
fig = go.Figure()
# Liste per creare una sola voce in legenda per Average e Best
avg_x, avg_y = [], []
best_x, best_y, best_text = [], [], []
for task in tasks:
avg_col = f"{task} Prompt Average"
best_col = f"{task} Best Prompt"
best_id_col = f"{task} Best Prompt Id"
if all(col in dataframe.columns for col in [avg_col, best_col, best_id_col]):
avg_value = dataframe[avg_col].mean()
avg_x.append(task)
avg_y.append(avg_value)
best_value = dataframe[best_col].mean()
best_x.append(task)
best_y.append(best_value)
best_id = dataframe[best_id_col].mode()[0] # Most frequent best prompt id
best_text.append(f"P:{best_id}")
# Barre Average Accuracy (azzurro)
fig.add_trace(go.Bar(
x=avg_x,
y=avg_y,
name="Avg. Accuracy",
marker_color="#1f77b4",
#hovertemplate="%{y:.2f}%
Accuracy: %{y:.2f}%
Accuracy: %{y:.2f}%
"
"Different prompts achieve the highest accuracy on different tasks.",
xref="paper", yref="paper",
x=0.5, y=-0.3,
showarrow=False,
font=dict(size=11, color="gray"),
align="center",
xanchor="center"
)
return fig
def line_chart(dataframe):
# Normalizza le dimensioni per avere marker non troppo piccoli né enormi
def scale_sizes(values, min_size=8, max_size=30):
vmin, vmax = min(values), max(values)
return [
min_size + (val - vmin) / (vmax - vmin) * (max_size - min_size) if vmax > vmin else (min_size + max_size) / 2
for val in values
]
# dati in base a IS_FS
df_true = dataframe[dataframe['IS_FS'] == True]
df_false = dataframe[dataframe['IS_FS'] == False]
# Estrai valori x, y e labels
x_true = df_true['#Params (B)'].tolist()
y_true = df_true['Avg. Comb. Perf. ⬆️'].tolist()
labels_true = [re.search(r'>([^<]+)<', m).group(1) for m in df_true['Model'].tolist()]
x_false = df_false['#Params (B)'].tolist()
y_false = df_false['Avg. Comb. Perf. ⬆️'].tolist()
labels_false = [re.search(r'>([^<]+)<', m).group(1) for m in df_false['Model'].tolist()]
fig = go.Figure()
# Punti IS_FS=True
fig.add_trace(go.Scatter(
x=x_true,
y=y_true,
mode='markers',
name='5-Shot',
marker=dict(
color='blue',
size=scale_sizes(x_true)
),
hovertemplate='%{customdata}
#Params: %{x}
Performance: %{y}
#Params: %{x}
Performance: %{y}
"
"with 5-shot can outperform larger zero-shot models.",
xref="paper", yref="paper",
x=0.5, y=-0.3, # 👈 centrata
showarrow=False,
font=dict(size=11, color="gray"),
align="center",
xanchor="center" # 👈 ancora centrata rispetto al testo
)
fig.update_xaxes(fixedrange=True, rangeslider_visible=False)
fig.update_yaxes(fixedrange=True)
return fig
# Define task metadata (icons, names, descriptions)
TASK_METADATA_MULTIPLECHOICE = {
"TE": {"icon": "📊", "name": "Textual Entailment", "tooltip": ""},
"SA": {"icon": "😃", "name": "Sentiment Analysis", "tooltip": ""},
"HS": {"icon": "⚠️", "name": "Hate Speech", "tooltip": ""},
"AT": {"icon": "🏥", "name": "Admission Test", "tooltip": ""},
"WIC": {"icon": "🔤", "name": "Word in Context", "tooltip": ""},
"FAQ": {"icon": "❓", "name": "Frequently Asked Questions", "tooltip": ""}
}
# Define task metadata (icons, names, descriptions)
TASK_METADATA_GENERATIVE = {
"LS": {"icon": "🔄", "name": "Lexical Substitution", "tooltip": ""},
"SU": {"icon": "📝", "name": "Summarization", "tooltip": ""},
"NER": {"icon": "🏷️", "name": "Named Entity Recognition", "tooltip": ""},
"REL": {"icon": "🔗", "name": "Relation Extraction", "tooltip": ""},
}
def restart_space():
"""Restart the Hugging Face space."""
API.restart_space(repo_id=REPO_ID)
def init_leaderboard(dataframe, default_selection=None, hidden_columns=None):
"""
Initialize and return the leaderboard when it is first loaded or when 'benchmark' is selected.
The table is sorted based on the "Avg. Combined Performance" field.
"""
if dataframe is None or dataframe.empty:
raise ValueError("Leaderboard DataFrame is empty or None.")
#print("????????????????????????????????", mean_of_max_per_field(dataframe))
sorted_dataframe = dataframe.sort_values(by="Avg. Comb. Perf. ⬆️", ascending=False)
sorted_dataframe = sorted_dataframe.reset_index(drop=True)
sorted_dataframe["Rank"] = sorted_dataframe.index + 1
# Flag per sapere se la medaglia è già stata assegnata per categoria e tipo
large_medal_fs_assigned = False
medium_medal_fs_assigned = False
small_medal_fs_assigned = False
large_medal_0shot_assigned = False
medium_medal_0shot_assigned = False
small_medal_0shot_assigned = False
# Lista temporanea per salvare i nuovi valori della colonna Model
new_model_column = []
for _, row in sorted_dataframe.iterrows():
if row['IS_FS']: # 5-Few-Shot
if row["Size"] == "🔵🔵🔵" and not large_medal_fs_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🔵🏆")
large_medal_fs_assigned = True
elif row["Size"] == "🔵🔵" and not medium_medal_fs_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🏆")
medium_medal_fs_assigned = True
elif row["Size"] == "🔵" and not small_medal_fs_assigned:
new_model_column.append(f"{row['Model']} 🔵🏆")
small_medal_fs_assigned = True
else:
new_model_column.append(row["Model"])
else: # 0-Shot
if row["Size"] == "🔵🔵🔵" and not large_medal_0shot_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🔵🎖️")
large_medal_0shot_assigned = True
elif row["Size"] == "🔵🔵" and not medium_medal_0shot_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🎖️")
medium_medal_0shot_assigned = True
elif row["Size"] == "🔵" and not small_medal_0shot_assigned:
new_model_column.append(f"{row['Model']} 🔵🎖️")
small_medal_0shot_assigned = True
else:
new_model_column.append(row["Model"])
# Lista delle colonne da aggiornare
#cols_to_update = ["REL Best Prompt Id", "NER Best Prompt Id", "SU Best Prompt Id", "LS Best Prompt Id"]
# Applichiamo la trasformazione
#for col in cols_to_update:
# dataframe[col] = dataframe[col].replace({1: 7, 2: 8})
# Aggiorna la colonna Model
sorted_dataframe["Model"] = new_model_column
field_list = fields(AutoEvalColumn)
return Leaderboard(
value=sorted_dataframe,
datatype=[c.type for c in field_list],
#select_columns=SelectColumns(
# default_selection=default_selection or [c.name for c in field_list if c.displayed_by_default],
# cant_deselect=[c.name for c in field_list if c.never_hidden],
# label="Select Columns to Display:",
#),
search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
filter_columns=[
ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
#ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Few-Shot Learning (FS)",
# default=[["0️⃣", "0️⃣"]]),
ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max = 100, default = [0,100], label="Select the number of parameters (B)"),
],
#filter_columns=[
# ColumnFilter("IS_FS", type="checkbox", default=False, label="5-Few-Shot")
# #ColumnFilter("FS", type="dropdown", label="5-Few-Shot")
#],
bool_checkboxgroup_label="Evaluation Mode",
interactive=False,
)
def update_task_leaderboard(dataframe, default_selection=None, hidden_columns=None):
"""
Update and return the leaderboard when a specific task is selected.
The table is sorted based on the "Combined Performance" field.
"""
if dataframe is None or dataframe.empty:
raise ValueError("Leaderboard DataFrame is empty or None.")
sorted_dataframe = dataframe.sort_values(by="Combined Performance", ascending=False)
# aggiungo la colonna rank in base alla posizione
sorted_dataframe = sorted_dataframe.reset_index(drop=True)
sorted_dataframe["Rank"] = sorted_dataframe.index + 1
# Flag per sapere se la medaglia è già stata assegnata per categoria e tipo
large_medal_fs_assigned = False
medium_medal_fs_assigned = False
small_medal_fs_assigned = False
large_medal_0shot_assigned = False
medium_medal_0shot_assigned = False
small_medal_0shot_assigned = False
# Lista temporanea per salvare i nuovi valori della colonna Model
new_model_column = []
for _, row in sorted_dataframe.iterrows():
if row['IS_FS']: # 5-Few-Shot
if row["Size"] == "🔵🔵🔵" and not large_medal_fs_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🔵🏆")
large_medal_fs_assigned = True
elif row["Size"] == "🔵🔵" and not medium_medal_fs_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🏆")
medium_medal_fs_assigned = True
elif row["Size"] == "🔵" and not small_medal_fs_assigned:
new_model_column.append(f"{row['Model']} 🔵🏆")
small_medal_fs_assigned = True
else:
new_model_column.append(row["Model"])
else: # 0-Shot
if row["Size"] == "🔵🔵🔵" and not large_medal_0shot_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🔵🎖️")
large_medal_0shot_assigned = True
elif row["Size"] == "🔵🔵" and not medium_medal_0shot_assigned:
new_model_column.append(f"{row['Model']} 🔵🔵🎖️")
medium_medal_0shot_assigned = True
elif row["Size"] == "🔵" and not small_medal_0shot_assigned:
new_model_column.append(f"{row['Model']} 🔵🎖️")
small_medal_0shot_assigned = True
else:
new_model_column.append(row["Model"])
# Aggiorna la colonna Model
sorted_dataframe["Model"] = new_model_column
pd.set_option('display.max_colwidth', None)
#print("========================", dataframe['Model'])
#print(sorted_dataframe['Combined Performance'])
field_list = fields(AutoEvalColumn)
return Leaderboard(
value=sorted_dataframe,
#datatype=[c.type for c in field_list],
datatype=[c.type for c in field_list] + [int],
#select_columns=SelectColumns(
# default_selection=default_selection or [c.name for c in field_list if c.displayed_by_default],
# cant_deselect=[c.name for c in field_list if c.never_hidden],
# label="Select Columns to Display:",
#),
search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
hide_columns=hidden_columns or [c.name for c in field_list if c.hidden],
filter_columns=[
ColumnFilter(AutoEvalColumn.fewshot_symbol.name, type="checkboxgroup", label="N-Shot Learning (FS)"),
ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max=100, default=[0, 100],
label="Select the number of parameters (B)"),
],
bool_checkboxgroup_label="Evaluation Mode",
interactive=False
)
'''
# Helper function for leaderboard initialization
def init_leaderboard(dataframe, default_selection=None, hidden_columns=None):
"""Initialize and return a leaderboard."""
if dataframe is None or dataframe.empty:
raise ValueError("Leaderboard DataFrame is empty or None.")
return Leaderboard(
value=dataframe,
datatype=[c.type for c in fields(AutoEvalColumn)],
select_columns=SelectColumns(
default_selection=default_selection or [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden],
label="Select Columns to Display:",
),
search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
hide_columns=hidden_columns or [c.name for c in fields(AutoEvalColumn) if c.hidden],
filter_columns=[
ColumnFilter(AutoEvalColumn.fewshot_type.name, type="checkboxgroup", label="N-Few-Shot Learning (FS)"),
ColumnFilter(AutoEvalColumn.params.name, type="slider", min=0, max=150, label="Select the number of parameters (B)"),
],
bool_checkboxgroup_label="Hide models",
interactive=False,
)
'''
def download_snapshot(repo, local_dir):
"""Try to download a snapshot from Hugging Face Hub."""
try:
print(f"Downloading from {repo} to {local_dir}...")
snapshot_download(repo_id=repo, local_dir=local_dir, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN)
except Exception as e:
print(f"Error downloading {repo}: {e}")
restart_space()
# Initialize the app by downloading snapshots
download_snapshot(QUEUE_REPO, EVAL_REQUESTS_PATH)
download_snapshot(RESULTS_REPO, EVAL_RESULTS_PATH)
# Load leaderboard data
LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
#print(LEADERBOARD_DF.columns.tolist())
theoretical_max_combined_perf = mean_of_max_per_field(LEADERBOARD_DF)
# Prepare the main interface
demo = gr.Blocks(css=custom_css)
with demo:
#gr.HTML(TITLE)
gr.HTML(
"""