Spaces:
Running
Running
File size: 9,125 Bytes
7c73423 b0c41e0 299dc31 7c73423 1706767 7c73423 1706767 0bba168 1706767 0bba168 1706767 0bba168 7c73423 e9d62ec 7c73423 47e1616 7c73423 47e1616 0bba168 7c73423 1706767 7c73423 47e1616 b0c41e0 efd894c b0c41e0 47e1616 7c73423 |
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 |
"""
TODO:
- 统计 tokenizer_impl
- 统计 OOV
- 统计 reversal
- 增加 math，code
## balance
- 高压缩率 VS vocab_size:
- 高压缩率，就意味着，编码后的token数少，那么 token长度 就会长，--> vocab_size 就会太大
- 高压缩率 VS 无损
- s
- OOV
- OOV 多，那么生成的 UNK 可能多(一个char 一个UNK) --> token 数目多 -> 压缩率低
- OOV 多，那么生成的 UNK 可能少（） --> token 数目多 -> 压缩率低
"""
import gradio as gr
from compression_util import get_compression_leaderboard, common_corpuses
# From the perspective of compression
# exactly reconstructed from compressed tokens
docs = """## 📖 What is a good tokenizer?
From a compression perspective, a good tokenizer should be lossless,
and keep high compression rate (fewer tokens for given text). <br>
The encoding and decoding process can be formulated as
```python
token_ids = tokenizer.encode(input_text) # compressed tokens
decoded_text = tokenizer.decode(token_ids) # reconstructed text
```
- **Lossless** <br>
Lossless tokenization preserves the exact original text, i.e. `decoded_text = input_text`. There are mainly two causes of compression loss.
1. `OOV`: Most lossy tokenizers get many out-of-vocabulary(OOV) words. 👉 Check the OOV and
tokenization loss of [bert](https://huggingface.co/spaces/eson/tokenizer-arena/blob/main/stats/compression_rate/google-bert.bert-base-cased%20%40%20cc100.zh-Hans.diff.json) and
[t5](https://huggingface.co/spaces/eson/tokenizer-arena/blob/main/stats/compression_rate/google-t5.t5-large%20%40%20cc100.es.diff.json).
2. `Normalization`: Even if a tokenizer has no OOV, it can be lossy due to text normalization. For example, qwen performs [unicode normalization](https://github.com/huggingface/transformers/blob/v4.42.3/src/transformers/models/qwen2/tokenization_qwen2.py#L338) in encoding process,
llama performs [clean_up_tokenization_spaces](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/main/tokenizer_config.json#L2053) in decoding process,
which may bring some slight differences to the reconstructed text. 👉 Check the tokenization loss of
[qwen](https://huggingface.co/spaces/eson/tokenizer-arena/raw/main/stats/compression_rate/Qwen.Qwen1.5-1.8B%20@%20cc100.ja.diff.json) and
[llama](https://huggingface.co/spaces/eson/tokenizer-arena/raw/main/stats/compression_rate/meta-llama.Meta-Llama-3.1-405B%20@%20cc100.en.diff.json).
- **Compression Rate** <br>
There are mainly two types of metric to represent the `input_text`:
- `char-level`: the number of characters in the given text
- `byte-level`: the number of bytes in the given text.
To evaluate compression rate, simple metrics can be "how many chars per token" or "how many bytes per token". <br>
In this leaderboard, we adopt more frequently used metric: "how many chars per token" and
"how many billion tokens per gigabytes corpus", i.e. `char/token` and `b_tokens/g_bytes`.
💬 [Discussion is Welcome](https://huggingface.co/spaces/eson/tokenizer-arena/discussions)
"""
# theme = gr.themes.Monochrome()
theme = gr.themes.Default()
# theme.set(accordion_text_weight=600) # 暂不支持
with gr.Blocks(theme=theme) as demo:
# gr.Markdown("## Convertor")
# with gr.Accordion("Convertor", open=False):
# gr.Markdown("Tokenize {} corpus")
# with gr.Row(elem_classes="no-border"):
# gr.Button("File Size", min_width=50)
# file_size = gr.Textbox(
# show_label=False,
# min_width=50,
# # elem_classes="textbox-as-text"
# )
# gr.Dropdown(
# choices=['MB', 'GB', 'TB'],
# show_label=False,
# min_width=15,
# # elem_classes="textbox-as-text"
# )
# # gr.Markdown('<h2 align="center">≈</h2>')
# # gr.HTML('<h2 style="margin: auto;">≈</h2>')
# gr.Button(
# "≈",
# min_width=10,
# elem_classes="button-white h2-font"
#
# )
#
# gr.Button(
# "Tokens",
# min_width=50
# )
# gr.Textbox(
# show_label=False,
# min_width=50
# )
# gr.Dropdown(
# ['million', 'billion', 'trillion'],
# show_label=False,
# min_width=15,
# elem_classes="button-white"
# )
gr.Markdown(docs)
gr.Markdown("## 🛠️ Setting") # ⚙
gr.Markdown("We perform tokenization on different corpus, and calculate the compression rate."
"")
with gr.Accordion("Please select the corpus and measure of compression rate.", open=True):
# file size 💽 🖴, tokens 🧮
# Total amount of disk used
with gr.Row():
with gr.Column():
compress_rate_corpus = gr.Dropdown(
common_corpuses, # , "code"
value=["cc100/en", "cc100/zh-Hans", "cc100/fr", "cc100/es"],
label="corpus",
multiselect=True
# info=""
)
# unit of file_size: gigabyte terabyte
# unit of token_num: million billion trillion
# The most common units of measurement include length (meter, inch, foot), weight (gram, kilogram, pound), volume (liter, gallon, milliliter), time (second, minute, hour)
compress_rate_unit = gr.Radio(
["b_tokens/g_bytes", "t_tokens/t_bytes"],
value="b_tokens/g_bytes",
label="measure", # evaluation metric
)
gr.Markdown(
# "Note:\n\n explanation"
# "Supported languages are (20): arabic (ar), bulgarian (bg), german (de), modern greek (el), english (en), spanish (es), french (fr), hindi (hi), italian (it), japanese (ja), dutch (nl), polish (pl), portuguese (pt), russian (ru), swahili (sw), thai (th), turkish (tr), urdu (ur), vietnamese (vi), and chinese (zh)."
# " arabic (ar), english (en), spanish (es), french (fr), italian (it), japanese (ja), portuguese (pt), russian (ru), and chinese (zh)."
"- `corpus`: tokenization is performed on the selected subsets of [cc100](https://huggingface.co/datasets/statmt/cc100) corpus.\n"
"- measure\n"
" - `b_tokens/g_bytes` measures how many billion tokens per gigabytes corpus.\n"
" - `t_tokens/t_bytes` measures how many trillion tokens per terabytes corpus.\n"
# "- `g_bytes/b_tokens` measures how many gigabytes corpus per billion tokens.\n"
# "- `t_bytes/t_tokens` measures how many terabytes corpus per trillion tokens.\n"
" - `char/token` measures how many chars per token on the tokenized corpus.\n"
" - `oov_ratio`: out-of-vocabulary ratio on the selected corpus, 👉 check [OOV charset](https://huggingface.co/spaces/eson/tokenizer-arena/raw/main/stats/compression_rate.json)\n\n"
"You can reproduce this procedure with [compression_util.py](https://huggingface.co/spaces/eson/tokenizer-arena/blob/main/compression_util.py)."
)
gr.Markdown("## 🏆 Compression Rate Leaderboard\n"
"This leaderboard aims to evaluate tokenizer performance on different languages.\n"
"Lower `oov_ratio` refers to fewer out-of-vocabulary tokens.\n"
"Lower `char/token` means more words might be segmented into subwords."
)
search_bar = gr.Textbox(
placeholder="🔍 Search by tokenizer or organization (e.g., 'llama', 'openai') and press ENTER...",
show_label=False,
elem_id="search-bar",
)
compress_rate_table = gr.Dataframe(datatype="html")
# func call
compress_rate_corpus.change(
get_compression_leaderboard,
inputs=[compress_rate_corpus, compress_rate_unit, search_bar],
outputs=compress_rate_table
)
compress_rate_unit.change(
get_compression_leaderboard,
inputs=[compress_rate_corpus, compress_rate_unit, search_bar],
outputs=compress_rate_table
)
# file_size.change(
# get_all_compress_rate,
# outputs=compress_rate_table
# )
search_bar.submit(
get_compression_leaderboard,
inputs=[
compress_rate_corpus,
compress_rate_unit,
search_bar,
],
outputs=compress_rate_table
)
demo.load(
get_compression_leaderboard,
inputs=[compress_rate_corpus, compress_rate_unit],
outputs=compress_rate_table
)
if __name__ == "__main__":
demo.launch()
|