File size: 8,298 Bytes
203ba9b 62d4a12 203ba9b 62d4a12 203ba9b 62d4a12 203ba9b 62d4a12 203ba9b 62d4a12 203ba9b 6a52aab 203ba9b 6a52aab 203ba9b 62d4a12 203ba9b 6a52aab 203ba9b a9aadc2 203ba9b 62d4a12 a9aadc2 62d4a12 203ba9b 62d4a12 203ba9b e4cac44 203ba9b 62d4a12 203ba9b 62d4a12 203ba9b 6a52aab 203ba9b 62d4a12 203ba9b a9aadc2 203ba9b 6a52aab a9aadc2 203ba9b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 |
from transformers import AutoTokenizer
from tqdm import tqdm
import gradio as gr
import pandas as pd
from datasets import load_dataset
import random
from pathlib import Path
initial_list_of_models = [
"Xenova/gpt-4o",
"NousResearch/Meta-Llama-3-8B",
"CohereForAI/c4ai-command-r-v01",
"CohereForAI/c4ai-command-r-plus",
"core42/jais-13b",
]
dataset = load_dataset("MohamedRashad/rasaif-translations", split="train")["arabic"]
dataframe_path = Path(__file__).parent / "arabic_tokenizers_leaderboard.jsonl"
if dataframe_path.exists():
df = pd.read_json(dataframe_path, lines=True)
else:
df = pd.DataFrame(columns=["📛 Models", "➕ Total Number of Tokens", "📘 Vocab Size", "Tokenizer Class"])
for model_name in tqdm(initial_list_of_models):
if model_name in df["📛 Models"].values:
continue
tokenizer = AutoTokenizer.from_pretrained(
model_name, use_fast=True, trust_remote_code=True
)
vocab_size = tokenizer.vocab_size
number_of_tokens = sum([len(x) for x in tokenizer(dataset).input_ids])
df = df._append(
{
"📛 Models": model_name,
"📘 Vocab Size": vocab_size,
"➕ Total Number of Tokens": number_of_tokens,
"Tokenizer Class": tokenizer.__class__.__name__,
},
ignore_index=True,
)
# Sort the dataframe by the number of tokens
df = df.sort_values(by="➕ Total Number of Tokens", ascending=True)
# Save the dataframe to a csv file
df.to_json(dataframe_path, lines=True, orient="records", force_ascii=False)
def submit(model_name):
global df
if model_name in df["📛 Models"].values:
return gr.Dataframe(df), gr.BarPlot(df), gr.Dropdown(choices=df["📛 Models"].tolist())
tokenizer = AutoTokenizer.from_pretrained(
model_name, use_fast=True, trust_remote_code=True
)
vocab_size = tokenizer.vocab_size
number_of_tokens = sum([len(x) for x in tokenizer(dataset).input_ids])
df = df._append(
{
"📛 Models": model_name,
"➕ Total Number of Tokens": number_of_tokens,
"📘 Vocab Size": vocab_size,
"Tokenizer Class": tokenizer.__class__.__name__,
},
ignore_index=True,
)
df = df.sort_values(by="➕ Total Number of Tokens", ascending=True)
df.to_json(dataframe_path, lines=True, orient="records", force_ascii=False)
return gr.Dataframe(df), gr.BarPlot(df), gr.Dropdown(choices=df["📛 Models"].tolist())
def generate_distinct_colors(n):
"""Generate n visually distinct colors in hexadecimal format."""
if n > 256**3:
raise ValueError("Cannot generate more than 16,777,216 unique colors.")
# To ensure colors are distinct, calculate an appropriate distance between colors
# The cube root of 256**3 (total colors) divided by n gives a crude initial spacing estimate
spacing = int((256 * 256 * 256)**(1/3) / n**(1/3))
max_val = 256 - spacing
# Set to keep track of used colors
used_colors = set()
# List to store the result colors
result = []
attempts = 0
while len(result) < n:
# Generate a color with a random start and controlled spacing
r = random.randint(0, max_val)
g = random.randint(0, max_val)
b = random.randint(0, max_val)
# Scale up by spacing to ensure minimum distance between colors
r = min(255, r * spacing)
g = min(255, g * spacing)
b = min(255, b * spacing)
# Format the color in hexadecimal
color = f"#{r:02X}{g:02X}{b:02X}"
# Ensure this color hasn't been used
if color not in used_colors:
used_colors.add(color)
result.append(color)
else:
attempts += 1
if attempts > 50:
# Dynamically adjust spacing if stuck
spacing = max(1, spacing - 1)
max_val = 256 - spacing
attempts = 0
return result
def decode_bpe_tokens(tokens):
fixed_tokens = []
for token in tokens:
# Check if the token starts with the special BPE space character 'Ġ'
if token.startswith('Ġ'):
# Process the rest of the token
try:
# Decode the rest of the token from UTF-8 bytes understood as Latin-1 characters
fixed_token = ' ' + token[1:].encode('utf-8').decode('utf-8')
except UnicodeDecodeError:
fixed_token = token # Use the original token if decoding fails
else:
try:
# Directly encode and decode without misinterpretation steps
fixed_token = token.encode('utf-8').decode('utf-8')
except UnicodeDecodeError:
fixed_token = token # Use the original token if decoding fails
fixed_tokens.append(fixed_token)
return fixed_tokens
def tokenize_text(text, chosen_model, better_tokenization=False):
tokenizer = AutoTokenizer.from_pretrained(chosen_model)
tokenized_text = decode_bpe_tokens(tokenizer.tokenize(text))
random_colors = generate_distinct_colors(len(tokenized_text))
if better_tokenization:
final_tokenized_text = []
for token in tokenized_text:
correct_tokenized_text = ""
for char in text:
correct_tokenized_text += char
current_token = decode_bpe_tokens(tokenizer.tokenize(correct_tokenized_text))
if current_token[0] == token:
final_tokenized_text.append(correct_tokenized_text)
text = text[len(correct_tokenized_text):]
break
else:
final_tokenized_text = tokenized_text
print(final_tokenized_text)
output = []
color_map = {}
for idx, token in enumerate(final_tokenized_text):
output.append((token, str(idx)))
color_map[str(idx+1)] = random_colors[idx % len(random_colors)]
return gr.HighlightedText(output, color_map)
leaderboard_description = """The `Total Number of Tokens` in this leaderboard is based on the total number of tokens summed on the Arabic section of [rasaif-translations](https://huggingface.co/datasets/MohamedRashad/rasaif-translations) dataset.
This dataset was chosen because it represents Arabic Fusha text in a small and consentrated manner.
A tokenizer that scores high in this leaderboard will be efficient in parsing Arabic in its different dialects and forms.
"""
with gr.Blocks() as demo:
gr.HTML("<center><h1>Arabic Tokenizers Leaderboard</h1></center>")
gr.Markdown("## What is the best tokenizer for Arabic?")
gr.Markdown(leaderboard_description)
with gr.Tab(label="Leaderboard"):
dataframe = gr.Dataframe(df)
with gr.Accordion("Barplot", open=False):
barplot = gr.BarPlot(
df,
x="📛 Models",
y="➕ Total Number of Tokens",
x_title=" ",
y_title=" ",
width=1000,
height=400,
tooltip=["📘 Vocab Size", "➕ Total Number of Tokens"],
vertical=False,
x_label_angle=30,
)
model_name = gr.Textbox(
label="Model Name from Hugging Face (e.g. Xenova/gpt-4o)"
)
submit_new_model_btn = gr.Button(value="Submit", variant="primary")
with gr.Tab(label="Try tokenizers"):
text = gr.Textbox(label="Enter a text", lines=5, value="السلام عليكم ورحمة الله", rtl=True, text_align="right")
dropdown = gr.Dropdown(
label="Select a model",
choices=df["📛 Models"].tolist(),
value=df["📛 Models"].tolist()[0],
)
with gr.Row():
submit_text_btn = gr.Button(value="Submit", variant="primary", scale=3)
checkbox = gr.Checkbox(label="Better tokenization for Arabic Text", value=False, scale=1)
tokenized_textbox = gr.HighlightedText(label="Tokenized text")
submit_new_model_btn.click(submit, model_name, outputs=[dataframe, barplot, dropdown])
submit_text_btn.click(tokenize_text, inputs=[text, dropdown, checkbox], outputs=[tokenized_textbox])
demo.launch()
|