QiYuanTokenizer-Large
QiYuanTokenizer is a universal multilingual tokenizer primarily optimized for Chinese–English mixed text,
offering compact and efficient tokenization across diverse languages and scripts.
It is designed as a general-purpose tokenizer, not tied to any specific model family,
and is especially suitable for encoder and encoder-decoder architectures.
✨ Overview
| Property | Value |
|---|
| Name | QiYuanTokenizer-Large |
| Type | Tokenizer-only repository |
| Purpose | General multilingual tokenization |
| Primary Languages | Chinese, English |
| Extended Support | Multilingual (Unicode-complete) |
| Architecture | Unigram |
| Vocabulary Size | 64,000 tokens |
| Fast Implementation | ✅ Available (QiYuanTokenizerFast) |
| Framework | 🤗 transformers |
| License | Apache 2.0 |
🧩 QiYuan Tokenizer Series
| Variant | Vocabulary Size | Description | Recommended Use |
|---|
| QiYuanTokenizer-Tiny | 12k | Extremely compact vocabulary for highly constrained settings. Efficient, but may become limiting for more demanding multilingual scenarios. | Use with caution |
| QiYuanTokenizer-Small | 24k | A lightweight tokenizer with improved coverage over Tiny while still keeping vocabulary size modest. | Compact models and efficiency-oriented experiments |
| QiYuanTokenizer-Base | 32k | A balanced baseline vocabulary suitable for general bilingual and multilingual tokenization tasks. | Recommended for general use |
| QiYuanTokenizer-Medium | 48k | The best-balanced variant in the series, providing strong coverage and good compression while keeping model complexity reasonable. | Recommended balance choice |
| QiYuanTokenizer-Large | 64k | A larger vocabulary designed for quality-oriented training, offering better coverage and stronger tokenization fidelity. | Recommended when quality is prioritized |
All variants share the same core token definitions and compatible special token settings.
⚙️ Usage
You can load this tokenizer directly with AutoTokenizer:
1from transformers import AutoTokenizer
2
3tokenizer = AutoTokenizer.from_pretrained("Morton-Li/QiYuanTokenizer-Large", trust_remote_code=True)
4
5# Example
6text = "你好，QiYuan！"
7tokens = tokenizer(text)
8print(tokens["input_ids"])
➕ Batch Example
1from transformers import AutoTokenizer
2
3tokenizer = AutoTokenizer.from_pretrained("Morton-Li/QiYuanTokenizer-Large", trust_remote_code=True)
4
5# Example
6texts = [
7 "Hello, 世界！",
8 "QiYuanTokenizer is designed for multilingual tokenization."
9]
10batch_tokens = tokenizer(texts, padding=True, return_tensors="pt")
11print(batch_tokens["input_ids"])
🧠 Design Notes
QiYuanTokenizer adopts the Unigram algorithm and is intended as a practical tokenizer for general text understanding and sequence transformation tasks.
In practice, it is generally more suitable for:
- Encoder models, such as text classification, embedding, retrieval, and sequence labeling
- Encoder-decoder models, such as translation, summarization, and text transformation
It can still be used in broader settings, but its design is not primarily oriented toward chat-format tokenization or decoder-only conversational templates.
📦 Files Included
| File | Description |
|---|
tokenizer.json | Serialized fast tokenizer definition |
tokenizer_config.json | Configuration (max length, padding side, etc.) |
tokenizer.py | Tokenizer implementation |
tokenizer.model | SentencePiece model file trained with the Unigram algorithm |
tokenizer.vocab | SentencePiece vocabulary file corresponding to tokenizer.model |
🔍 Special Tokens
| Token | Purpose |
|---|
<|unk|> | Unknown token |
<|bos|> | Beginning of sequence |
<|eos|> | End of sequence |
<|pad|> | Padding token for batch alignment |
<|mask|> | Masked token for MLM-style objectives |
🔖 License
This tokenizer and vocabulary are released under the Apache License 2.0.
You are free to use, modify, and redistribute it under the same license terms.
📚 Citation
If you use QiYuanTokenizer in your research or project, please cite it as:
1@misc{QiYuanTokenizer,
2 title = {QiYuanTokenizer: A Universal Multilingual Unigram Tokenizer with Chinese-English Optimization},
3 author = {Morton Li},
4 year = {2026},
5}
