Lyte's picture
Update app.py
c0bbbe3 verified
import gradio as gr
from transformers import AutoTokenizer
import pandas as pd
import re
from datetime import datetime
from huggingface_hub import HfApi, DatasetCard, DatasetCardData, create_repo
from gradio_huggingfacehub_search import HuggingfaceHubSearch
import os
import tempfile
import re
# --- Configuration ---
HF_TOKEN = os.getenv("HF_TOKEN")
DATASET_REPO_ID = os.getenv("DATASET_REPO", "Lyte/tokenizer-leaderboard")
DATASET_FILE_NAME = "leaderboard.csv"
PREDEFINED_TEXT = '''
import gradio as gr
from transformers import AutoTokenizer
import pandas as pd
import re
from datetime import datetime
from huggingface_hub import HfApi, DatasetCard, DatasetCardData, create_repo
from gradio_huggingfacehub_search import HuggingfaceHubSearch
import os
import tempfile
# --- Configuration ---
HF_TOKEN = os.getenv("HF_TOKEN")
DATASET_REPO_ID = os.getenv("DATASET_REPO", "Lyte/tokenizer-leaderboard")
DATASET_FILE_NAME = "leaderboard.csv"
PREDEFINED_TEXT = """
The quick brown fox jumps over 12 lazy dogs! 🐕‍🦺
Special characters: #@%^&*()_+-=[]{}|;:'",.<>/?\\~
Code samples:
- Python: def hello(): print("Hello World! 2023")
- HTML: <div class="container" id="main">Content</div>
- JSON: {"key": "value", "numbers": [1, 2, 3.14]}
Math equations: E = mc² → 3×10⁸ m/s
Multilingual text: 速い茶色の狐が怠惰な犬を飛び越える 😸
Emojis: 👍🎉🚀❤️🔥
Mixed casing: OpenAI's GPT-4 vs gpt-3.5-turbo
"""
WORD_COUNT = len(re.findall(r'\S+', PREDEFINED_TEXT))
LEADERBOARD_COLUMNS = [
"Model ID", "Token Count", "Vocab Size",
"Tokens/Word", "Chars/Token", "Timestamp"
]
# --- Hugging Face Hub Functions ---
def create_huggingface_dataset():
"""Creates the dataset repository on the Hub if it doesn't exist."""
try:
api = HfApi(token=HF_TOKEN)
create_repo(repo_id=DATASET_REPO_ID, token=HF_TOKEN, repo_type="dataset", exist_ok=True)
card_data = DatasetCardData(
language="en",
license="mit",
size_categories=["1K<n<10K"],
tags=["tokenizer", "leaderboard", "performance", "gradio"],
)
card = DatasetCard.from_template(
card_data,
template_path=None,
Title="Tokenizer Leaderboard",
Description="A leaderboard of tokenizer performance based on various metrics.",
How_to_use="The leaderboard data is stored in a CSV file named 'leaderboard.csv'.",
)
card.push_to_hub(repo_id=DATASET_REPO_ID, token=HF_TOKEN)
print(f"Dataset repository '{DATASET_REPO_ID}' created (or already exists).")
except Exception as e:
print(f"Error creating dataset repository: {e}")
raise
def load_leaderboard_from_hub():
"""Loads the leaderboard data from the Hugging Face Hub as a pandas DataFrame."""
try:
api = HfApi(token=HF_TOKEN)
dataset_path = api.dataset_info(repo_id=DATASET_REPO_ID, token=HF_TOKEN).siblings
csv_file_info = next((file for file in dataset_path if file.rfilename == DATASET_FILE_NAME), None)
if csv_file_info is None:
print(f"'{DATASET_FILE_NAME}' not found in '{DATASET_REPO_ID}'. Returning an empty DataFrame")
return pd.DataFrame(columns=LEADERBOARD_COLUMNS)
file_path = api.hf_hub_download(repo_id=DATASET_REPO_ID, filename=DATASET_FILE_NAME, repo_type="dataset")
df = pd.read_csv(file_path)
df = df.sort_values(by="Token Count", ascending=True)
df["Tokens/Word"] = df["Tokens/Word"].round(2)
df["Chars/Token"] = df["Chars/Token"].round(2)
return df
except Exception as e:
print(f"Error loading leaderboard from Hugging Face Hub: {e}")
return pd.DataFrame(columns=LEADERBOARD_COLUMNS)
def push_leaderboard_to_hub(df):
"""Pushes the updated leaderboard DataFrame to the Hugging Face Hub."""
try:
with tempfile.NamedTemporaryFile(mode='w', delete=False, suffix=".csv") as tmpfile:
df.to_csv(tmpfile.name, index=False)
tmp_path = tmpfile.name
api = HfApi(token=HF_TOKEN)
api.upload_file(
path_or_fileobj=tmp_path,
path_in_repo=DATASET_FILE_NAME,
repo_id=DATASET_REPO_ID,
repo_type="dataset",
token=HF_TOKEN,
commit_message="Update leaderboard"
)
os.remove(tmp_path)
print(f"Leaderboard updated and pushed to {DATASET_REPO_ID}")
except Exception as e:
print(f"Error pushing leaderboard to Hugging Face Hub: {e}")
raise
# --- Utility Functions ---
def get_tokenizer_stats(model_id, text):
if not model_id:
raise ValueError("No model ID provided")
try:
tokenizer = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN, trust_remote_code=True)
tokens = tokenizer.encode(text, add_special_tokens=False)
text_length = len(text)
return {
"token_count": len(tokens),
"vocab_size": tokenizer.vocab_size,
"token_word_ratio": round(len(tokens) / WORD_COUNT, 2),
"chars_per_token": round(text_length / len(tokens), 2) if tokens else 0
}
except Exception as e:
raise RuntimeError(f"Failed to load tokenizer or encode text: {str(e)}") from e
def is_model_in_leaderboard(df, model_id):
return model_id in df["Model ID"].values
def add_to_leaderboard(model_id):
if not model_id:
return "❌ Error: No model ID provided"
df = load_leaderboard_from_hub()
if is_model_in_leaderboard(df, model_id):
return "⚠️ Model already in leaderboard"
try:
stats = get_tokenizer_stats(model_id, PREDEFINED_TEXT)
new_row = pd.DataFrame([{
"Model ID": model_id,
"Token Count": stats["token_count"],
"Vocab Size": stats["vocab_size"],
"Tokens/Word": stats["token_word_ratio"],
"Chars/Token": stats["chars_per_token"],
"Timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S")
}])
updated_df = pd.concat([df, new_row], ignore_index=True)
push_leaderboard_to_hub(updated_df)
return "✅ Added to leaderboard!"
except Exception as e:
return f"❌ Error: {str(e)}"
def analyze_tokenizer(model_id, text):
if not model_id:
return "❌ Error: Please select or enter a model ID"
try:
stats = get_tokenizer_stats(model_id, text)
return (
f"Token Count: {stats['token_count']}\n"
f"Vocab Size: {stats['vocab_size']}\n"
f"Tokens/Word: {stats['token_word_ratio']:.2f}\n"
f"Chars/Token: {stats['chars_per_token']:.2f}"
)
except Exception as e:
return f"❌ Analysis Failed: {str(e)}"
def compare_tokenizers(model_ids_str, use_standard_text):
try:
model_list = [mid.strip() for mid in model_ids_str.split(',') if mid.strip()]
if not model_list:
return pd.DataFrame({"Error": ["No models provided"]})
results = []
for model_id in model_list:
try:
stats = get_tokenizer_stats(model_id, PREDEFINED_TEXT)
results.append({
"Model ID": model_id,
"Tokens": stats["token_count"],
"Vocab Size": stats["vocab_size"],
"Tokens/Word": f"{stats['token_word_ratio']:.2f}",
"Chars/Token": f"{stats['chars_per_token']:.2f}",
"Status": "✅ Success"
})
except Exception as e:
results.append({
"Model ID": model_id,
"Tokens": "-",
"Vocab Size": "-",
"Tokens/Word": "-",
"Chars/Token": "-",
"Status": f"❌ {str(e)}"
})
return pd.DataFrame(results)
except Exception as e:
return pd.DataFrame({"Error": [str(e)]})
def get_leaderboard_for_download():
"""Loads, prepares, and returns a Gradio File object for download."""
try:
df = load_leaderboard_from_hub()
with tempfile.NamedTemporaryFile(delete=False, suffix=".csv") as tmpfile:
df.to_csv(tmpfile.name, index=False)
# Return a Gradio File object, NOT just the path
return gr.File(value=tmpfile.name, label="Download CSV")
except Exception as e:
print(f"Error preparing file for download: {e}")
return None
def initial_benchmark_run():
try:
print("Starting initial benchmark run...")
default_models = [
"deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
"Qwen/Qwen2.5-7B-Instruct-1M",
"simplescaling/s1.1-32B",
"Xenova/gpt-4o",
"microsoft/phi-4",
"deepseek-ai/DeepSeek-R1",
"google/gemma-2-27b-it",
"HuggingFaceTB/SmolLM2-135M-Instruct",
"mistralai/Mistral-7B-Instruct-v0.3",
"tomg-group-umd/huginn-0125",
"microsoft/Phi-3.5-mini-instruct",
"openai-community/gpt2"
]
df = load_leaderboard_from_hub()
for model_id in default_models:
try:
if not is_model_in_leaderboard(df, model_id):
print(f"Benchmarking {model_id}...")
result = add_to_leaderboard(model_id)
print(f"Result for {model_id}: {result}")
else:
print(f"{model_id} already in leaderboard, skipping.")
except Exception as e:
print(f"Error benchmarking {model_id}: {str(e)}")
print("Initial benchmarking complete.")
except Exception as e:
print(f"Fatal error in initial benchmark: {str(e)}")
# --- Gradio Interface ---
with gr.Blocks(title="Tokenizers Leaderboard", theme=gr.themes.Soft()) as iface:
gr.Markdown("# 🏆 Tokenizers Leaderboard")
with gr.Tab("Analyze"):
gr.Markdown("## Single Tokenizer Analysis")
with gr.Row():
model_search = HuggingfaceHubSearch(label="Search Models", placeholder="Search Hugging Face models...", search_type="model")
custom_model = gr.Textbox(label="Direct Model ID", placeholder="e.g.: mistralai/Mistral-7B-Instruct-v0.3", max_lines=1)
model_id = gr.Textbox(visible=False)
gr.Markdown("### Input Text")
text_input = gr.Textbox(lines=5, value=PREDEFINED_TEXT, label="Analysis Text")
with gr.Row():
analyze_btn = gr.Button("Analyze", variant="primary")
add_btn = gr.Button("Add to Leaderboard")
analysis_output = gr.Textbox(label="Results", interactive=False)
model_search.change(lambda x: x, model_search, model_id)
custom_model.change(lambda x: x, custom_model, model_id)
analyze_btn.click(analyze_tokenizer, [model_id, text_input], analysis_output)
add_event = add_btn.click(add_to_leaderboard, model_id, analysis_output)
with gr.Tab("Compare"):
gr.Markdown("## Multi-Model Comparison")
gr.Markdown(f"**Standard Text:** `{PREDEFINED_TEXT[:80]}...`")
model_ids = gr.Textbox(label="Model IDs (comma-separated)", placeholder="Enter models: meta-llama/Llama-2-7b, google/gemma-7b, ...")
compare_btn = gr.Button("Compare Models", variant="primary")
comparison_table = gr.DataFrame(label="Results", interactive=False)
compare_btn.click(compare_tokenizers, [model_ids, gr.Checkbox(value=True, visible=False)], comparison_table)
with gr.Tab("Leaderboard"):
gr.Markdown("## Performance Leaderboard")
with gr.Row():
download_btn = gr.DownloadButton(label="Download CSV", value="tokenizer_leaderboard.csv")
leaderboard_table = gr.DataFrame(label="Top Tokenizers", headers=LEADERBOARD_COLUMNS, interactive=False,
datatype=["str", "number", "number", "number", "number", "str"])
# Connect the download button to the function that prepares the CSV
download_btn.click(get_leaderboard_for_download, inputs=[], outputs=download_btn)
iface.load(fn=load_leaderboard_from_hub, outputs=leaderboard_table)
add_event.then(load_leaderboard_from_hub, None, leaderboard_table)
create_huggingface_dataset()
initial_benchmark_run()
iface.launch()
'''
WORD_COUNT = len(re.findall(r'\S+', PREDEFINED_TEXT))
LEADERBOARD_COLUMNS = [
"Model ID", "Token Count", "Vocab Size",
"Tokens/Word", "Chars/Token", "Timestamp"
]
# --- Hugging Face Hub Functions ---
def create_huggingface_dataset():
"""Creates the dataset repository on the Hub if it doesn't exist."""
try:
api = HfApi(token=HF_TOKEN)
create_repo(repo_id=DATASET_REPO_ID, token=HF_TOKEN, repo_type="dataset", exist_ok=True)
card_data = DatasetCardData(
language="en",
license="mit",
size_categories=["1K<n<10K"],
tags=["tokenizer", "leaderboard", "performance", "gradio"],
)
card = DatasetCard.from_template(
card_data,
template_path=None,
Title="Tokenizer Leaderboard",
Description="A leaderboard of tokenizer performance based on various metrics.",
How_to_use="The leaderboard data is stored in a CSV file named 'leaderboard.csv'.",
)
card.push_to_hub(repo_id=DATASET_REPO_ID, token=HF_TOKEN)
print(f"Dataset repository '{DATASET_REPO_ID}' created (or already exists).")
except Exception as e:
print(f"Error creating dataset repository: {e}")
raise
def load_leaderboard_from_hub():
"""Loads the leaderboard data from the Hugging Face Hub as a pandas DataFrame."""
try:
api = HfApi(token=HF_TOKEN)
dataset_path = api.dataset_info(repo_id=DATASET_REPO_ID, token=HF_TOKEN).siblings
csv_file_info = next((file for file in dataset_path if file.rfilename == DATASET_FILE_NAME), None)
if csv_file_info is None:
print(f"'{DATASET_FILE_NAME}' not found in '{DATASET_REPO_ID}'. Returning an empty DataFrame")
return pd.DataFrame(columns=LEADERBOARD_COLUMNS)
file_path = api.hf_hub_download(repo_id=DATASET_REPO_ID, filename=DATASET_FILE_NAME, repo_type="dataset")
df = pd.read_csv(file_path)
df = df.sort_values(by="Token Count", ascending=True)
df["Tokens/Word"] = df["Tokens/Word"].round(2)
df["Chars/Token"] = df["Chars/Token"].round(2)
return df
except Exception as e:
print(f"Error loading leaderboard from Hugging Face Hub: {e}")
return pd.DataFrame(columns=LEADERBOARD_COLUMNS)
def push_leaderboard_to_hub(df):
"""Pushes the updated leaderboard DataFrame to the Hugging Face Hub."""
try:
with tempfile.NamedTemporaryFile(mode='w', delete=False, suffix=".csv") as tmpfile:
df.to_csv(tmpfile.name, index=False)
tmp_path = tmpfile.name
api = HfApi(token=HF_TOKEN)
api.upload_file(
path_or_fileobj=tmp_path,
path_in_repo=DATASET_FILE_NAME,
repo_id=DATASET_REPO_ID,
repo_type="dataset",
token=HF_TOKEN,
commit_message="Update leaderboard"
)
os.remove(tmp_path)
print(f"Leaderboard updated and pushed to {DATASET_REPO_ID}")
except Exception as e:
print(f"Error pushing leaderboard to Hugging Face Hub: {e}")
raise
# --- Utility Functions ---
def get_tokenizer_stats(model_id, text):
if not model_id:
raise ValueError("No model ID provided")
try:
tokenizer = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN, trust_remote_code=True)
tokens = tokenizer.encode(text, add_special_tokens=False)
text_length = len(text)
return {
"token_count": len(tokens),
"vocab_size": tokenizer.vocab_size,
"token_word_ratio": round(len(tokens) / WORD_COUNT, 2),
"chars_per_token": round(text_length / len(tokens), 2) if tokens else 0
}
except Exception as e:
raise RuntimeError(f"Failed to load tokenizer or encode text: {str(e)}") from e
def is_model_in_leaderboard(df, model_id):
return model_id in df["Model ID"].values
def add_to_leaderboard(model_id):
if not model_id:
return "❌ Error: No model ID provided"
df = load_leaderboard_from_hub()
if is_model_in_leaderboard(df, model_id):
return "⚠️ Model already in leaderboard"
try:
stats = get_tokenizer_stats(model_id, PREDEFINED_TEXT)
new_row = pd.DataFrame([{
"Model ID": model_id,
"Token Count": stats["token_count"],
"Vocab Size": stats["vocab_size"],
"Tokens/Word": stats["token_word_ratio"],
"Chars/Token": stats["chars_per_token"],
"Timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S")
}])
updated_df = pd.concat([df, new_row], ignore_index=True)
push_leaderboard_to_hub(updated_df)
return "✅ Added to leaderboard!"
except Exception as e:
return f"❌ Error: {str(e)}"
def analyze_tokenizer(model_id, text):
if not model_id:
return "❌ Error: Please select or enter a model ID"
try:
stats = get_tokenizer_stats(model_id, text)
return (
f"Token Count: {stats['token_count']}\n"
f"Vocab Size: {stats['vocab_size']}\n"
f"Tokens/Word: {stats['token_word_ratio']:.2f}\n"
f"Chars/Token: {stats['chars_per_token']:.2f}"
)
except Exception as e:
return f"❌ Analysis Failed: {str(e)}"
def compare_tokenizers(model_ids_str, use_standard_text):
try:
model_list = [mid.strip() for mid in model_ids_str.split(',') if mid.strip()]
if not model_list:
return pd.DataFrame({"Error": ["No models provided"]})
results = []
for model_id in model_list:
try:
stats = get_tokenizer_stats(model_id, PREDEFINED_TEXT)
results.append({
"Model ID": model_id,
"Tokens": stats["token_count"],
"Vocab Size": stats["vocab_size"],
"Tokens/Word": f"{stats['token_word_ratio']:.2f}",
"Chars/Token": f"{stats['chars_per_token']:.2f}",
"Status": "✅ Success"
})
except Exception as e:
results.append({
"Model ID": model_id,
"Tokens": "-",
"Vocab Size": "-",
"Tokens/Word": "-",
"Chars/Token": "-",
"Status": f"❌ {str(e)}"
})
return pd.DataFrame(results)
except Exception as e:
return pd.DataFrame({"Error": [str(e)]})
def get_leaderboard_for_download():
"""Loads, prepares, and returns a Gradio File object for download."""
try:
df = load_leaderboard_from_hub()
with tempfile.NamedTemporaryFile(delete=False, suffix=".csv") as tmpfile:
df.to_csv(tmpfile.name, index=False)
# Return a Gradio File object, NOT just the path
return gr.File(value=tmpfile.name, label="Download CSV")
except Exception as e:
print(f"Error preparing file for download: {e}")
return None
def initial_benchmark_run():
try:
print("Starting initial benchmark run...")
default_models = [
"deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
"Qwen/Qwen2.5-7B-Instruct-1M",
"simplescaling/s1.1-32B",
"Xenova/gpt-4o",
"microsoft/phi-4",
"deepseek-ai/DeepSeek-R1",
"google/gemma-2-27b-it",
"HuggingFaceTB/SmolLM2-135M-Instruct",
"mistralai/Mistral-7B-Instruct-v0.3",
"tomg-group-umd/huginn-0125",
"microsoft/Phi-3.5-mini-instruct",
"openai-community/gpt2"
]
df = load_leaderboard_from_hub()
for model_id in default_models:
try:
if not is_model_in_leaderboard(df, model_id):
print(f"Benchmarking {model_id}...")
result = add_to_leaderboard(model_id)
print(f"Result for {model_id}: {result}")
else:
print(f"{model_id} already in leaderboard, skipping.")
except Exception as e:
print(f"Error benchmarking {model_id}: {str(e)}")
print("Initial benchmarking complete.")
except Exception as e:
print(f"Fatal error in initial benchmark: {str(e)}")
# --- Gradio Interface ---
with gr.Blocks(title="Tokenizers Leaderboard", theme=gr.themes.Soft()) as iface:
gr.Markdown("# 🏆 Tokenizers Leaderboard")
with gr.Tab("Analyze"):
gr.Markdown("## Single Tokenizer Analysis")
with gr.Row():
model_search = HuggingfaceHubSearch(label="Search Models", placeholder="Search Hugging Face models...", search_type="model")
custom_model = gr.Textbox(label="Direct Model ID", placeholder="e.g.: mistralai/Mistral-7B-Instruct-v0.3", max_lines=1)
model_id = gr.Textbox(visible=False)
gr.Markdown("### Input Text")
text_input = gr.Textbox(lines=5, value=PREDEFINED_TEXT, label="Analysis Text")
with gr.Row():
analyze_btn = gr.Button("Analyze", variant="primary")
add_btn = gr.Button("Add to Leaderboard")
analysis_output = gr.Textbox(label="Results", interactive=False)
model_search.change(lambda x: x, model_search, model_id)
custom_model.change(lambda x: x, custom_model, model_id)
analyze_btn.click(analyze_tokenizer, [model_id, text_input], analysis_output)
add_event = add_btn.click(add_to_leaderboard, model_id, analysis_output)
with gr.Tab("Compare"):
gr.Markdown("## Multi-Model Comparison")
gr.Markdown(f"**Standard Text:** `{PREDEFINED_TEXT[:80]}...`")
model_ids = gr.Textbox(label="Model IDs (comma-separated)", placeholder="Enter models: meta-llama/Llama-2-7b, google/gemma-7b, ...")
compare_btn = gr.Button("Compare Models", variant="primary")
comparison_table = gr.DataFrame(label="Results", interactive=False)
compare_btn.click(compare_tokenizers, [model_ids, gr.Checkbox(value=True, visible=False)], comparison_table)
with gr.Tab("Leaderboard"):
gr.Markdown("## Performance Leaderboard")
gr.Markdown(f"The tokenizers are run on a predefined text of {len(PREDEFINED_TEXT)} Length which has a word count of {WORD_COUNT}")
with gr.Row():
download_btn = gr.DownloadButton(label="Download CSV", value="tokenizer_leaderboard.csv")
leaderboard_table = gr.DataFrame(label="Top Tokenizers", headers=LEADERBOARD_COLUMNS, interactive=False,
datatype=["str", "number", "number", "number", "number", "str"])
# Connect the download button to the function that prepares the CSV
download_btn.click(get_leaderboard_for_download, inputs=[], outputs=download_btn)
iface.load(fn=load_leaderboard_from_hub, outputs=leaderboard_table)
add_event.then(load_leaderboard_from_hub, None, leaderboard_table)
create_huggingface_dataset()
initial_benchmark_run()
iface.launch()