Spaces:
Running
Running
metadata
title: Tokenizer Arena
emoji: ⚡
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 3.41.2
app_file: app.py
pinned: false
压缩率 Compress Rate
在 cc-100 数据集,每个语言取1万条数据,测试不同tokenizer的压缩率。压缩率指标 g_bytes/b_tokens
您可通过以下脚本进行复现
python utils/compress_rate_util.py
简体中文压缩率
在简体中文数据集 cc100-zh-Hans 计算压缩率tokenizer | vocab_size | g_bytes/b_tokens | t_bytes/t_tokens | b_tokens/g_bytes |
---|---|---|---|---|
amber | 32000 | 1.84 | 1.8 | 0.54 |
aya_101 | 250100 | 3.89 | 3.79 | 0.26 |
baichuan | 64000 | 3.92 | 3.82 | 0.26 |
baichuan2 | 125696 | 4.53 | 4.42 | 0.22 |
bert_base_cased | 28996 | 2.73 | 2.66 | 0.37 |
bert_base_chinese | 21128 | 2.74 | 2.67 | 0.37 |
bert_base_uncased | 30522 | 2.73 | 2.67 | 0.37 |
bloom | 250680 | 4.28 | 4.18 | 0.23 |
byt5_small | 256 | 0.93 | 0.91 | 1.08 |
character_glm_6b | 64794 | 4.2 | 4.1 | 0.24 |
chatglm2_6b | 64794 | 4.2 | 4.1 | 0.24 |
chatglm3_6b | 64798 | 4.2 | 4.1 | 0.24 |
chatglm_6b | 150344 | 4.65 | 4.54 | 0.22 |
chatyuan_large_v2 | 32128 | 4.34 | 4.24 | 0.23 |
chinese_llama | 49953 | 3.93 | 3.84 | 0.25 |
chinese_llama2 | 55296 | 3.92 | 3.83 | 0.26 |
code_davinci_002 | 50281 | 1.31 | 1.28 | 0.77 |
crystal_coder | 32000 | 1.86 | 1.81 | 0.54 |
deepseek_coder_33b_instruct | 32000 | 3.4 | 3.32 | 0.29 |
deepseek_llm_7b_base | 100000 | 4.05 | 3.96 | 0.25 |
falcon_180b | 65024 | 2.18 | 2.13 | 0.46 |
falcon_7b | 65024 | 2.18 | 2.13 | 0.46 |
fastchat_t5_3b | 32000 | 13.7 | 13.38 | 0.07 |
flan_t5_base | 32100 | 14.13 | 13.8 | 0.07 |
gemma_7b | 256000 | 3.82 | 3.73 | 0.26 |
gpt2 | 50257 | 1.31 | 1.28 | 0.77 |
gpt2_chinese | 21128 | 2.73 | 2.66 | 0.37 |
gpt_35_turbo | 100277 | 2.26 | 2.21 | 0.44 |
gpt_4 | 100277 | 2.26 | 2.21 | 0.44 |
gpt_nexo_20b | 50254 | 2.01 | 1.96 | 0.5 |
internlm2_chat_7b | 92544 | 4.23 | 4.13 | 0.24 |
internlm2_math_7b | 92544 | 4.23 | 4.13 | 0.24 |
internlm_chat_7b | 103168 | 4.23 | 4.14 | 0.24 |
internlm_xcomposer_7b | 103168 | 4.23 | 4.14 | 0.24 |
kplug | 10261 | 2.72 | 2.65 | 0.37 |
llama | 32000 | 1.84 | 1.8 | 0.54 |
llama2 | 32000 | 1.84 | 1.8 | 0.54 |
mistral_7b | 32000 | 2.36 | 2.3 | 0.42 |
mixtral_8_7b | 32000 | 2.36 | 2.3 | 0.42 |
mobilebert_uncased | 30522 | 2.73 | 2.67 | 0.37 |
moss | 106029 | 4.4 | 4.3 | 0.23 |
mt5_large | 250100 | 3.89 | 3.79 | 0.26 |
olmo_7b | 50280 | 2.01 | 1.96 | 0.5 |
orion_14b_chat | 84608 | 4.63 | 4.52 | 0.22 |
phi_1 | 50257 | 1.31 | 1.28 | 0.77 |
phi_2 | 50257 | 1.31 | 1.28 | 0.77 |
pko_t5_large | 50258 | 0.97 | 0.95 | 1.03 |
prompt_clue | 32128 | 4.34 | 4.24 | 0.23 |
qwen1_5_14b_chat | 151643 | 4.16 | 4.06 | 0.24 |
qwen_1_8b_chat | 151851 | 4.16 | 4.06 | 0.24 |
qwen_72b_chat | 151851 | 4.16 | 4.06 | 0.24 |
qwen_7b_chat | 151851 | 4.16 | 4.06 | 0.24 |
roberta_chinese_clue | 8021 | 2.7 | 2.64 | 0.37 |
skywork_13b_base | 65519 | 3.69 | 3.61 | 0.27 |
skywork_13b_math | 65519 | 3.69 | 3.61 | 0.27 |
solar_10_7b | 32000 | 2.36 | 2.3 | 0.42 |
starchat_alpha | 49152 | 2.78 | 2.72 | 0.36 |
switch_c_2048 | 32100 | 14.13 | 13.8 | 0.07 |
t5_base | 32100 | 14.13 | 13.8 | 0.07 |
t5_large | 32100 | 14.13 | 13.8 | 0.07 |
t5_small | 32100 | 14.13 | 13.8 | 0.07 |
text_davinci_003 | 50281 | 1.31 | 1.28 | 0.77 |
tigerbot_13b_chat_v2 | 60512 | 4.25 | 4.15 | 0.24 |
tigerbot_70b_chat_v4_4k | 65107 | 4.25 | 4.15 | 0.24 |
wizardcoder_15b_v1 | 49152 | 2.78 | 2.72 | 0.36 |
wizardcoder_python_7b_v1 | 32000 | 1.84 | 1.8 | 0.54 |
wizardlm_7b_v1 | 32000 | 1.84 | 1.8 | 0.54 |
wizardmath_70b_v1 | 32000 | 1.84 | 1.8 | 0.54 |
xlm_roberta | 250002 | 3.96 | 3.86 | 0.25 |
yi_34b | 64000 | 4.17 | 4.07 | 0.24 |
yi_6b | 64000 | 4.17 | 4.07 | 0.24 |
yi_vl34b | 64000 | 4.11 | 4.02 | 0.24 |
zephyr_7b_beta | 32000 | 2.36 | 2.3 | 0.42 |
结论 larger vocabulary sizes
Reference
- Getting the most out of your tokenizer for pre-training and domain adaptation
- Efficient and Effective Text Encoding for Chinese LLaMA and Alpaca
- https://huggingface.co/spaces/Xenova/the-tokenizer-playground