import gradio as gr
import pandas as pd
import os
import zipfile
import base64
CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""@misc{aienergyscore-leaderboard,
author = {Sasha Luccioni and Boris Gamazaychikov and Emma Strubell and Sara Hooker and Yacine Jernite and Carole-Jean Wu and Margaret Mitchell},
title = {AI Energy Score Leaderboard - February 2025},
year = {2025},
publisher = {Hugging Face},
howpublished = "\url{https://huggingface.co/spaces/AIEnergyScore/Leaderboard}",
}"""
# List of tasks (CSV filenames)
tasks = [
'asr.csv',
'object_detection.csv',
'text_classification.csv',
'image_captioning.csv',
'question_answering.csv',
'text_generation.csv',
'image_classification.csv',
'sentence_similarity.csv',
'image_generation.csv',
'summarization.csv'
]
### HELPER FUNCTIONS ###
def format_stars(score):
try:
score_int = int(score)
except Exception:
score_int = 0
return f'{"★" * score_int}'
def make_link(mname):
parts = str(mname).split('/')
display_name = parts[1] if len(parts) > 1 else mname
return f'{display_name}'
def extract_link_text(html_link):
start = html_link.find('>') + 1
end = html_link.rfind('')
if start > 0 and end > start:
return html_link[start:end]
else:
return html_link
def generate_html_table_from_df(df):
# Compute a static width for the Model column based on the longest model name.
if not df.empty:
max_length = max(len(extract_link_text(link)) for link in df['Model'])
else:
max_length = 10
static_width = max_length * 10 + 16
max_energy = df['gpu_energy_numeric'].max() if not df.empty else 1
color_map = {"1": "black", "2": "black", "3": "black", "4": "black", "5": "black"}
html = ''
html += ''
html += '| Model | '
html += 'Provider | '
html += 'GPU Energy (Wh) | '
html += 'Score | '
html += '
'
html += ''
for _, row in df.iterrows():
energy_numeric = row['gpu_energy_numeric']
# Format energy with commas and 2 decimal places.
energy_str = f"{energy_numeric:,.2f}"
bar_width = (energy_numeric / max_energy) * 100
score_val = row['energy_score']
bar_color = color_map.get(str(score_val), "gray")
html += ''
html += f'| {row["Model"]} | '
html += f'{row["Provider"]} | '
html += (
f'{energy_str}
'
f' | '
)
html += f'{row["Score"]} | '
html += '
'
html += '
'
return f'{html}
'
def process_df(task, sort_order="Low to High", filter_fn=None):
df = pd.read_csv('data/energy/' + task)
if df.columns[0].startswith("Unnamed:"):
df = df.iloc[:, 1:]
df['energy_score'] = df['energy_score'].astype(int)
df['gpu_energy_numeric'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000
if filter_fn is not None:
df = filter_fn(df)
df['Provider'] = df['model'].apply(lambda x: str(x).split('/')[0])
df['Model'] = df['model'].apply(make_link)
df['Score'] = df['energy_score'].apply(format_stars)
ascending = True if sort_order == "Low to High" else False
df = df.sort_values(by='gpu_energy_numeric', ascending=ascending)
return df
def compute_efficiency_ratio(df):
if df.empty:
return 1
min_val = df['gpu_energy_numeric'].min()
max_val = df['gpu_energy_numeric'].max()
ratio = max_val / min_val if min_val > 0 else 1
return ratio
def generate_callout(ratio, scope_text):
"""
Returns a right-aligned callout where the inner box is shrink-wrapped to its text.
The ratio is formatted with a comma for thousands.
"""
return (
f''
f'
'
f' Energy efficiency difference of {ratio:,.1f}x for {scope_text}.'
f'
'
f'
Select different tasks to see scored models.
")
with gr.Column(scale=4):
global_callout = gr.HTML(get_global_callout())
# --- Tabs for the different tasks ---
with gr.Tabs():
# --- Text Generation Tab ---
with gr.TabItem("Text Generation 💬"):
with gr.Row():
with gr.Column(scale=8):
model_class_options = [
"A (Single Consumer GPU) <20B parameters",
"B (Single Cloud GPU) 20-66B parameters",
"C (Multiple Cloud GPUs) >66B parameters"
]
model_class_dropdown = gr.Dropdown(
choices=model_class_options,
label="Select Model Class",
value=model_class_options[0]
)
sort_dropdown_tg = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
tg_callout = gr.HTML()
tg_table = gr.HTML()
init_callout, init_table = update_text_generation(model_class_options[0], "Low to High")
tg_callout.value = init_callout
tg_table.value = init_table
model_class_dropdown.change(fn=update_text_generation, inputs=[model_class_dropdown, sort_dropdown_tg], outputs=[tg_callout, tg_table])
sort_dropdown_tg.change(fn=update_text_generation, inputs=[model_class_dropdown, sort_dropdown_tg], outputs=[tg_callout, tg_table])
# --- Image Generation Tab ---
with gr.TabItem("Image Generation 📷"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_img = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
img_callout = gr.HTML()
img_table = gr.HTML()
init_callout, init_table = update_image_generation("Low to High")
img_callout.value = init_callout
img_table.value = init_table
sort_dropdown_img.change(fn=update_image_generation, inputs=sort_dropdown_img, outputs=[img_callout, img_table])
# --- Text Classification Tab ---
with gr.TabItem("Text Classification 🎭"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_tc = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
tc_callout = gr.HTML()
tc_table = gr.HTML()
init_callout, init_table = update_text_classification("Low to High")
tc_callout.value = init_callout
tc_table.value = init_table
sort_dropdown_tc.change(fn=update_text_classification, inputs=sort_dropdown_tc, outputs=[tc_callout, tc_table])
# --- Image Classification Tab ---
with gr.TabItem("Image Classification 🖼️"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_ic = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
ic_callout = gr.HTML()
ic_table = gr.HTML()
init_callout, init_table = update_image_classification("Low to High")
ic_callout.value = init_callout
ic_table.value = init_table
sort_dropdown_ic.change(fn=update_image_classification, inputs=sort_dropdown_ic, outputs=[ic_callout, ic_table])
# --- Image Captioning Tab ---
with gr.TabItem("Image Captioning 📝"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_icap = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
icap_callout = gr.HTML()
icap_table = gr.HTML()
init_callout, init_table = update_image_captioning("Low to High")
icap_callout.value = init_callout
icap_table.value = init_table
sort_dropdown_icap.change(fn=update_image_captioning, inputs=sort_dropdown_icap, outputs=[icap_callout, icap_table])
# --- Summarization Tab ---
with gr.TabItem("Summarization 📃"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_sum = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
sum_callout = gr.HTML()
sum_table = gr.HTML()
init_callout, init_table = update_summarization("Low to High")
sum_callout.value = init_callout
sum_table.value = init_table
sort_dropdown_sum.change(fn=update_summarization, inputs=sort_dropdown_sum, outputs=[sum_callout, sum_table])
# --- Automatic Speech Recognition Tab ---
with gr.TabItem("Automatic Speech Recognition 💬"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_asr = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
asr_callout = gr.HTML()
asr_table = gr.HTML()
init_callout, init_table = update_asr("Low to High")
asr_callout.value = init_callout
asr_table.value = init_table
sort_dropdown_asr.change(fn=update_asr, inputs=sort_dropdown_asr, outputs=[asr_callout, asr_table])
# --- Object Detection Tab ---
with gr.TabItem("Object Detection 🚘"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_od = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
od_callout = gr.HTML()
od_table = gr.HTML()
init_callout, init_table = update_object_detection("Low to High")
od_callout.value = init_callout
od_table.value = init_table
sort_dropdown_od.change(fn=update_object_detection, inputs=sort_dropdown_od, outputs=[od_callout, od_table])
# --- Sentence Similarity Tab ---
with gr.TabItem("Sentence Similarity 📚"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_ss = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
ss_callout = gr.HTML()
ss_table = gr.HTML()
init_callout, init_table = update_sentence_similarity("Low to High")
ss_callout.value = init_callout
ss_table.value = init_table
sort_dropdown_ss.change(fn=update_sentence_similarity, inputs=sort_dropdown_ss, outputs=[ss_callout, ss_table])
# --- Extractive QA Tab ---
with gr.TabItem("Extractive QA ❔"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_qa = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
qa_callout = gr.HTML()
qa_table = gr.HTML()
init_callout, init_table = update_extractive_qa("Low to High")
qa_callout.value = init_callout
qa_table.value = init_table
sort_dropdown_qa.change(fn=update_extractive_qa, inputs=sort_dropdown_qa, outputs=[qa_callout, qa_table])
# --- All Tasks Tab ---
with gr.TabItem("All Tasks 💡"):
with gr.Row():
with gr.Column(scale=8):
sort_dropdown_all = gr.Dropdown(
choices=["Low to High", "High to Low"],
label="Sort",
value="Low to High"
)
with gr.Column(scale=4):
all_callout = gr.HTML()
all_table = gr.HTML()
init_callout, init_table = update_all_tasks("Low to High")
all_callout.value = init_callout
all_table.value = init_table
sort_dropdown_all.change(fn=update_all_tasks, inputs=sort_dropdown_all, outputs=[all_callout, all_table])
with gr.Accordion("📙 Citation", open=False):
citation_button = gr.Textbox(
value=CITATION_BUTTON_TEXT,
label=CITATION_BUTTON_LABEL,
elem_id="citation-button",
lines=10,
show_copy_button=True,
)
gr.Markdown("Last updated: February 2025")
demo.launch() 