[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6G9HYDRCCAMJF9MME4MBW3B":3,"skill-same:01M6G9HYDRCCAMJF9MME4MBW3B":76},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":54,"commitSha":55,"skillDirectory":56,"forksCount":57,"repositoryHtmlUrl":58,"repositoryHomepageUrl":6,"ownerHtmlUrl":59,"ownerBlogUrl":6,"ownerDescription":60,"updatedAt":61,"install":62,"resources":75},"01M6G9HYDRCCAMJF9MME4MBW3B","huggingface-tokenizers",null,"Fast BPE\u002FWordPiece tokenization and custom vocab training.","zapabob\u002FSkills","zapabob","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F130830376?v=4","01M6DBMYZAGRTFQDSQYW3W764P","峯岸 亮","01M6DBMZ3GKDSJ01NZ9XTZD7JD",2,"MIT","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Farchive\u002Fskill-md-reviewed\u002Fsha256\u002F79\u002F79ca48477b7b72a9d5b59e472c89bec4399048ba2f5e9cc3698ed6e6facc5b0c\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"HuggingFace","huggingface tokenizers","alignment","Alignment Tracking","Custom Tokenizer","Fast Tokenization","Unigram","WordPiece","BPE","Tokenization","datasets","macos","Production","Transformers","fast","training","windows","research","rust","linux","2026-11-10 10:12:30",false,"---\nname: huggingface-tokenizers\ndescription: Fast BPE\u002FWordPiece tokenization and custom vocab training.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ndependencies: [tokenizers, transformers, datasets]\nplatforms: [linux, macos, windows]\nmetadata:\n  hermes:\n    tags: [Tokenization, HuggingFace, BPE, WordPiece, Unigram, Fast Tokenization, Rust, Custom Tokenizer, Alignment Tracking, Production]\n\n---\n\n# HuggingFace Tokenizers - Fast Tokenization for NLP\n\nFast, production-ready tokenizers with Rust performance and Python ease-of-use.\n\n## When to use HuggingFace Tokenizers\n\n**Use HuggingFace Tokenizers when:**\n- Need extremely fast tokenization (\u003C20s per GB of text)\n- Training custom tokenizers from scratch\n- Want alignment tracking (token → original text position)\n- Building production NLP pipelines\n- Need to tokenize large corpora efficiently\n\n**Performance**:\n- **Speed**: \u003C20 seconds to tokenize 1GB on CPU\n- **Implementation**: Rust core with Python\u002FNode.js bindings\n- **Efficiency**: 10-100× faster than pure Python implementations\n\n**Use alternatives instead**:\n- **SentencePiece**: Language-independent, used by T5\u002FALBERT\n- **tiktoken**: OpenAI's BPE tokenizer for GPT models\n- **transformers AutoTokenizer**: Loading pretrained only (uses this library internally)\n\n## Quick start\n\n### Installation\n\n```bash\n# Install tokenizers\npip install tokenizers\n\n# With transformers integration\npip install tokenizers transformers\n```\n\n### Load pretrained tokenizer\n\n```python\nfrom tokenizers import Tokenizer\n\n# Load from HuggingFace Hub\ntokenizer = Tokenizer.from_pretrained(\"bert-base-uncased\")\n\n# Encode text\noutput = tokenizer.encode(\"Hello, how are you?\")\nprint(output.tokens)  # ['hello', ',', 'how', 'are', 'you', '?']\nprint(output.ids)     # [7592, 1010, 2129, 2024, 2017, 1029]\n\n# Decode back\ntext = tokenizer.decode(output.ids)\nprint(text)  # \"hello, how are you?\"\n```\n\n### Train custom BPE tokenizer\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import BPE\nfrom tokenizers.trainers import BpeTrainer\nfrom tokenizers.pre_tokenizers import Whitespace\n\n# Initialize tokenizer with BPE model\ntokenizer = Tokenizer(BPE(unk_token=\"[UNK]\"))\ntokenizer.pre_tokenizer = Whitespace()\n\n# Configure trainer\ntrainer = BpeTrainer(\n    vocab_size=30000,\n    special_tokens=[\"[UNK]\", \"[CLS]\", \"[SEP]\", \"[PAD]\", \"[MASK]\"],\n    min_frequency=2\n)\n\n# Train on files\nfiles = [\"train.txt\", \"validation.txt\"]\ntokenizer.train(files, trainer)\n\n# Save\ntokenizer.save(\"my-tokenizer.json\")\n```\n\n**Training time**: ~1-2 minutes for 100MB corpus, ~10-20 minutes for 1GB\n\n### Batch encoding with padding\n\n```python\n# Enable padding\ntokenizer.enable_padding(pad_id=3, pad_token=\"[PAD]\")\n\n# Encode batch\ntexts = [\"Hello world\", \"This is a longer sentence\"]\nencodings = tokenizer.encode_batch(texts)\n\nfor encoding in encodings:\n    print(encoding.ids)\n# [101, 7592, 2088, 102, 3, 3, 3]\n# [101, 2023, 2003, 1037, 2936, 6251, 102]\n```\n\n## Tokenization algorithms\n\n### BPE (Byte-Pair Encoding)\n\n**How it works**:\n1. Start with character-level vocabulary\n2. Find most frequent character pair\n3. Merge into new token, add to vocabulary\n4. Repeat until vocabulary size reached\n\n**Used by**: GPT-2, GPT-3, RoBERTa, BART, DeBERTa\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import BPE\nfrom tokenizers.trainers import BpeTrainer\nfrom tokenizers.pre_tokenizers import ByteLevel\n\ntokenizer = Tokenizer(BPE(unk_token=\"\u003C|endoftext|>\"))\ntokenizer.pre_tokenizer = ByteLevel()\n\ntrainer = BpeTrainer(\n    vocab_size=50257,\n    special_tokens=[\"\u003C|endoftext|>\"],\n    min_frequency=2\n)\n\ntokenizer.train(files=[\"data.txt\"], trainer=trainer)\n```\n\n**Advantages**:\n- Handles OOV words well (breaks into subwords)\n- Flexible vocabulary size\n- Good for morphologically rich languages\n\n**Trade-offs**:\n- Tokenization depends on merge order\n- May split common words unexpectedly\n\n### WordPiece\n\n**How it works**:\n1. Start with character vocabulary\n2. Score merge pairs: `frequency(pair) \u002F (frequency(first) × frequency(second))`\n3. Merge highest scoring pair\n4. Repeat until vocabulary size reached\n\n**Used by**: BERT, DistilBERT, MobileBERT\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import WordPiece\nfrom tokenizers.trainers import WordPieceTrainer\nfrom tokenizers.pre_tokenizers import Whitespace\nfrom tokenizers.normalizers import BertNormalizer\n\ntokenizer = Tokenizer(WordPiece(unk_token=\"[UNK]\"))\ntokenizer.normalizer = BertNormalizer(lowercase=True)\ntokenizer.pre_tokenizer = Whitespace()\n\ntrainer = WordPieceTrainer(\n    vocab_size=30522,\n    special_tokens=[\"[UNK]\", \"[CLS]\", \"[SEP]\", \"[PAD]\", \"[MASK]\"],\n    continuing_subword_prefix=\"##\"\n)\n\ntokenizer.train(files=[\"corpus.txt\"], trainer=trainer)\n```\n\n**Advantages**:\n- Prioritizes meaningful merges (high score = semantically related)\n- Used successfully in BERT (state-of-the-art results)\n\n**Trade-offs**:\n- Unknown words become `[UNK]` if no subword match\n- Saves vocabulary, not merge rules (larger files)\n\n### Unigram\n\n**How it works**:\n1. Start with large vocabulary (all substrings)\n2. Compute loss for corpus with current vocabulary\n3. Remove tokens with minimal impact on loss\n4. Repeat until vocabulary size reached\n\n**Used by**: ALBERT, T5, mBART, XLNet (via SentencePiece)\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import Unigram\nfrom tokenizers.trainers import UnigramTrainer\n\ntokenizer = Tokenizer(Unigram())\n\ntrainer = UnigramTrainer(\n    vocab_size=8000,\n    special_tokens=[\"\u003Cunk>\", \"\u003Cs>\", \"\u003C\u002Fs>\"],\n    unk_token=\"\u003Cunk>\"\n)\n\ntokenizer.train(files=[\"data.txt\"], trainer=trainer)\n```\n\n**Advantages**:\n- Probabilistic (finds most likely tokenization)\n- Works well for languages without word boundaries\n- Handles diverse linguistic contexts\n\n**Trade-offs**:\n- Computationally expensive to train\n- More hyperparameters to tune\n\n## Tokenization pipeline\n\nComplete pipeline: **Normalization → Pre-tokenization → Model → Post-processing**\n\n### Normalization\n\nClean and standardize text:\n\n```python\nfrom tokenizers.normalizers import NFD, StripAccents, Lowercase, Sequence\n\ntokenizer.normalizer = Sequence([\n    NFD(),           # Unicode normalization (decompose)\n    Lowercase(),     # Convert to lowercase\n    StripAccents()   # Remove accents\n])\n\n# Input: \"Héllo WORLD\"\n# After normalization: \"hello world\"\n```\n\n**Common normalizers**:\n- `NFD`, `NFC`, `NFKD`, `NFKC` - Unicode normalization forms\n- `Lowercase()` - Convert to lowercase\n- `StripAccents()` - Remove accents (é → e)\n- `Strip()` - Remove whitespace\n- `Replace(pattern, content)` - Regex replacement\n\n### Pre-tokenization\n\nSplit text into word-like units:\n\n```python\nfrom tokenizers.pre_tokenizers import Whitespace, Punctuation, Sequence, ByteLevel\n\n# Split on whitespace and punctuation\ntokenizer.pre_tokenizer = Sequence([\n    Whitespace(),\n    Punctuation()\n])\n\n# Input: \"Hello, world!\"\n# After pre-tokenization: [\"Hello\", \",\", \"world\", \"!\"]\n```\n\n**Common pre-tokenizers**:\n- `Whitespace()` - Split on spaces, tabs, newlines\n- `ByteLevel()` - GPT-2 style byte-level splitting\n- `Punctuation()` - Isolate punctuation\n- `Digits(individual_digits=True)` - Split digits individually\n- `Metaspace()` - Replace spaces with ▁ (SentencePiece style)\n\n### Post-processing\n\nAdd special tokens for model input:\n\n```python\nfrom tokenizers.processors import TemplateProcessing\n\n# BERT-style: [CLS] sentence [SEP]\ntokenizer.post_processor = TemplateProcessing(\n    single=\"[CLS] $A [SEP]\",\n    pair=\"[CLS] $A [SEP] $B [SEP]\",\n    special_tokens=[\n        (\"[CLS]\", 1),\n        (\"[SEP]\", 2),\n    ],\n)\n```\n\n**Common patterns**:\n```python\n# GPT-2: sentence \u003C|endoftext|>\nTemplateProcessing(\n    single=\"$A \u003C|endoftext|>\",\n    special_tokens=[(\"\u003C|endoftext|>\", 50256)]\n)\n\n# RoBERTa: \u003Cs> sentence \u003C\u002Fs>\nTemplateProcessing(\n    single=\"\u003Cs> $A \u003C\u002Fs>\",\n    pair=\"\u003Cs> $A \u003C\u002Fs> \u003C\u002Fs> $B \u003C\u002Fs>\",\n    special_tokens=[(\"\u003Cs>\", 0), (\"\u003C\u002Fs>\", 2)]\n)\n```\n\n## Alignment tracking\n\nTrack token positions in original text:\n\n```python\noutput = tokenizer.encode(\"Hello, world!\")\n\n# Get token offsets\nfor token, offset in zip(output.tokens, output.offsets):\n    start, end = offset\n    print(f\"{token:10} → [{start:2}, {end:2}): {text[start:end]!r}\")\n\n# Output:\n# hello      → [ 0,  5): 'Hello'\n# ,          → [ 5,  6): ','\n# world      → [ 7, 12): 'world'\n# !          → [12, 13): '!'\n```\n\n**Use cases**:\n- Named entity recognition (map predictions back to text)\n- Question answering (extract answer spans)\n- Token classification (align labels to original positions)\n\n## Integration with transformers\n\n### Load with AutoTokenizer\n\n```python\nfrom transformers import AutoTokenizer\n\n# AutoTokenizer automatically uses fast tokenizers\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\n\n# Check if using fast tokenizer\nprint(tokenizer.is_fast)  # True\n\n# Access underlying tokenizers.Tokenizer\nfast_tokenizer = tokenizer.backend_tokenizer\nprint(type(fast_tokenizer))  # \u003Cclass 'tokenizers.Tokenizer'>\n```\n\n### Convert custom tokenizer to transformers\n\n```python\nfrom tokenizers import Tokenizer\nfrom transformers import PreTrainedTokenizerFast\n\n# Train custom tokenizer\ntokenizer = Tokenizer(BPE())\n# ... train tokenizer ...\ntokenizer.save(\"my-tokenizer.json\")\n\n# Wrap for transformers\ntransformers_tokenizer = PreTrainedTokenizerFast(\n    tokenizer_file=\"my-tokenizer.json\",\n    unk_token=\"[UNK]\",\n    pad_token=\"[PAD]\",\n    cls_token=\"[CLS]\",\n    sep_token=\"[SEP]\",\n    mask_token=\"[MASK]\"\n)\n\n# Use like any transformers tokenizer\noutputs = transformers_tokenizer(\n    \"Hello world\",\n    padding=True,\n    truncation=True,\n    max_length=512,\n    return_tensors=\"pt\"\n)\n```\n\n## Common patterns\n\n### Train from iterator (large datasets)\n\n```python\nfrom datasets import load_dataset\n\n# Load dataset\ndataset = load_dataset(\"wikitext\", \"wikitext-103-raw-v1\", split=\"train\")\n\n# Create batch iterator\ndef batch_iterator(batch_size=1000):\n    for i in range(0, len(dataset), batch_size):\n        yield dataset[i:i + batch_size][\"text\"]\n\n# Train tokenizer\ntokenizer.train_from_iterator(\n    batch_iterator(),\n    trainer=trainer,\n    length=len(dataset)  # For progress bar\n)\n```\n\n**Performance**: Processes 1GB in ~10-20 minutes\n\n### Enable truncation and padding\n\n```python\n# Enable truncation\ntokenizer.enable_truncation(max_length=512)\n\n# Enable padding\ntokenizer.enable_padding(\n    pad_id=tokenizer.token_to_id(\"[PAD]\"),\n    pad_token=\"[PAD]\",\n    length=512  # Fixed length, or None for batch max\n)\n\n# Encode with both\noutput = tokenizer.encode(\"This is a long sentence that will be truncated...\")\nprint(len(output.ids))  # 512\n```\n\n### Multi-processing\n\n```python\nfrom tokenizers import Tokenizer\nfrom multiprocessing import Pool\n\n# Load tokenizer\ntokenizer = Tokenizer.from_file(\"tokenizer.json\")\n\ndef encode_batch(texts):\n    return tokenizer.encode_batch(texts)\n\n# Process large corpus in parallel\nwith Pool(8) as pool:\n    # Split corpus into chunks\n    chunk_size = 1000\n    chunks = [corpus[i:i+chunk_size] for i in range(0, len(corpus), chunk_size)]\n\n    # Encode in parallel\n    results = pool.map(encode_batch, chunks)\n```\n\n**Speedup**: 5-8× with 8 cores\n\n## Performance benchmarks\n\n### Training speed\n\n| Corpus Size | BPE (30k vocab) | WordPiece (30k) | Unigram (8k) |\n|-------------|-----------------|-----------------|--------------|\n| 10 MB       | 15 sec          | 18 sec          | 25 sec       |\n| 100 MB      | 1.5 min         | 2 min           | 4 min        |\n| 1 GB        | 15 min          | 20 min          | 40 min       |\n\n**Hardware**: 16-core CPU, tested on English Wikipedia\n\n### Tokenization speed\n\n| Implementation | 1 GB corpus | Throughput    |\n|----------------|-------------|---------------|\n| Pure Python    | ~20 minutes | ~50 MB\u002Fmin    |\n| HF Tokenizers  | ~15 seconds | ~4 GB\u002Fmin     |\n| **Speedup**    | **80×**     | **80×**       |\n\n**Test**: English text, average sentence length 20 words\n\n### Memory usage\n\n| Task                    | Memory  |\n|-------------------------|---------|\n| Load tokenizer          | ~10 MB  |\n| Train BPE (30k vocab)   | ~200 MB |\n| Encode 1M sentences     | ~500 MB |\n\n## Supported models\n\nPre-trained tokenizers available via `from_pretrained()`:\n\n**BERT family**:\n- `bert-base-uncased`, `bert-large-cased`\n- `distilbert-base-uncased`\n- `roberta-base`, `roberta-large`\n\n**GPT family**:\n- `gpt2`, `gpt2-medium`, `gpt2-large`\n- `distilgpt2`\n\n**T5 family**:\n- `t5-small`, `t5-base`, `t5-large`\n- `google\u002Fflan-t5-xxl`\n\n**Other**:\n- `facebook\u002Fbart-base`, `facebook\u002Fmbart-large-cc25`\n- `albert-base-v2`, `albert-xlarge-v2`\n- `xlm-roberta-base`, `xlm-roberta-large`\n\nBrowse all: https:\u002F\u002Fhuggingface.co\u002Fmodels?library=tokenizers\n\n## References\n\n- **[Training Guide](references\u002Ftraining.md)** - Train custom tokenizers, configure trainers, handle large datasets\n- **[Algorithms Deep Dive](references\u002Falgorithms.md)** - BPE, WordPiece, Unigram explained in detail\n- **[Pipeline Components](references\u002Fpipeline.md)** - Normalizers, pre-tokenizers, post-processors, decoders\n- **[Transformers Integration](references\u002Fintegration.md)** - AutoTokenizer, PreTrainedTokenizerFast, special tokens\n\n## Resources\n\n- **Docs**: https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftokenizers\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftokenizers ⭐ 9,000+\n- **Version**: 0.20.0+\n- **Course**: https:\u002F\u002Fhuggingface.co\u002Flearn\u002Fnlp-course\u002Fchapter6\u002F1\n- **Paper**: BPE (Sennrich et al., 2016), WordPiece (Schuster & Nakajima, 2012)\n","# HuggingFace Tokenizers - Fast Tokenization for NLP\n\nFast, production-ready tokenizers with Rust performance and Python ease-of-use.\n\n## When to use HuggingFace Tokenizers\n\n**Use HuggingFace Tokenizers when:**\n- Need extremely fast tokenization (\u003C20s per GB of text)\n- Training custom tokenizers from scratch\n- Want alignment tracking (token → original text position)\n- Building production NLP pipelines\n- Need to tokenize large corpora efficiently\n\n**Performance**:\n- **Speed**: \u003C20 seconds to tokenize 1GB on CPU\n- **Implementation**: Rust core with Python\u002FNode.js bindings\n- **Efficiency**: 10-100× faster than pure Python implementations\n\n**Use alternatives instead**:\n- **SentencePiece**: Language-independent, used by T5\u002FALBERT\n- **tiktoken**: OpenAI's BPE tokenizer for GPT models\n- **transformers AutoTokenizer**: Loading pretrained only (uses this library internally)\n\n## Quick start\n\n### Installation\n\n```bash\n# Install tokenizers\npip install tokenizers\n\n# With transformers integration\npip install tokenizers transformers\n```\n\n### Load pretrained tokenizer\n\n```python\nfrom tokenizers import Tokenizer\n\n# Load from HuggingFace Hub\ntokenizer = Tokenizer.from_pretrained(\"bert-base-uncased\")\n\n# Encode text\noutput = tokenizer.encode(\"Hello, how are you?\")\nprint(output.tokens)  # ['hello', ',', 'how', 'are', 'you', '?']\nprint(output.ids)     # [7592, 1010, 2129, 2024, 2017, 1029]\n\n# Decode back\ntext = tokenizer.decode(output.ids)\nprint(text)  # \"hello, how are you?\"\n```\n\n### Train custom BPE tokenizer\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import BPE\nfrom tokenizers.trainers import BpeTrainer\nfrom tokenizers.pre_tokenizers import Whitespace\n\n# Initialize tokenizer with BPE model\ntokenizer = Tokenizer(BPE(unk_token=\"[UNK]\"))\ntokenizer.pre_tokenizer = Whitespace()\n\n# Configure trainer\ntrainer = BpeTrainer(\n    vocab_size=30000,\n    special_tokens=[\"[UNK]\", \"[CLS]\", \"[SEP]\", \"[PAD]\", \"[MASK]\"],\n    min_frequency=2\n)\n\n# Train on files\nfiles = [\"train.txt\", \"validation.txt\"]\ntokenizer.train(files, trainer)\n\n# Save\ntokenizer.save(\"my-tokenizer.json\")\n```\n\n**Training time**: ~1-2 minutes for 100MB corpus, ~10-20 minutes for 1GB\n\n### Batch encoding with padding\n\n```python\n# Enable padding\ntokenizer.enable_padding(pad_id=3, pad_token=\"[PAD]\")\n\n# Encode batch\ntexts = [\"Hello world\", \"This is a longer sentence\"]\nencodings = tokenizer.encode_batch(texts)\n\nfor encoding in encodings:\n    print(encoding.ids)\n# [101, 7592, 2088, 102, 3, 3, 3]\n# [101, 2023, 2003, 1037, 2936, 6251, 102]\n```\n\n## Tokenization algorithms\n\n### BPE (Byte-Pair Encoding)\n\n**How it works**:\n1. Start with character-level vocabulary\n2. Find most frequent character pair\n3. Merge into new token, add to vocabulary\n4. Repeat until vocabulary size reached\n\n**Used by**: GPT-2, GPT-3, RoBERTa, BART, DeBERTa\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import BPE\nfrom tokenizers.trainers import BpeTrainer\nfrom tokenizers.pre_tokenizers import ByteLevel\n\ntokenizer = Tokenizer(BPE(unk_token=\"\u003C|endoftext|>\"))\ntokenizer.pre_tokenizer = ByteLevel()\n\ntrainer = BpeTrainer(\n    vocab_size=50257,\n    special_tokens=[\"\u003C|endoftext|>\"],\n    min_frequency=2\n)\n\ntokenizer.train(files=[\"data.txt\"], trainer=trainer)\n```\n\n**Advantages**:\n- Handles OOV words well (breaks into subwords)\n- Flexible vocabulary size\n- Good for morphologically rich languages\n\n**Trade-offs**:\n- Tokenization depends on merge order\n- May split common words unexpectedly\n\n### WordPiece\n\n**How it works**:\n1. Start with character vocabulary\n2. Score merge pairs: `frequency(pair) \u002F (frequency(first) × frequency(second))`\n3. Merge highest scoring pair\n4. Repeat until vocabulary size reached\n\n**Used by**: BERT, DistilBERT, MobileBERT\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import WordPiece\nfrom tokenizers.trainers import WordPieceTrainer\nfrom tokenizers.pre_tokenizers import Whitespace\nfrom tokenizers.normalizers import BertNormalizer\n\ntokenizer = Tokenizer(WordPiece(unk_token=\"[UNK]\"))\ntokenizer.normalizer = BertNormalizer(lowercase=True)\ntokenizer.pre_tokenizer = Whitespace()\n\ntrainer = WordPieceTrainer(\n    vocab_size=30522,\n    special_tokens=[\"[UNK]\", \"[CLS]\", \"[SEP]\", \"[PAD]\", \"[MASK]\"],\n    continuing_subword_prefix=\"##\"\n)\n\ntokenizer.train(files=[\"corpus.txt\"], trainer=trainer)\n```\n\n**Advantages**:\n- Prioritizes meaningful merges (high score = semantically related)\n- Used successfully in BERT (state-of-the-art results)\n\n**Trade-offs**:\n- Unknown words become `[UNK]` if no subword match\n- Saves vocabulary, not merge rules (larger files)\n\n### Unigram\n\n**How it works**:\n1. Start with large vocabulary (all substrings)\n2. Compute loss for corpus with current vocabulary\n3. Remove tokens with minimal impact on loss\n4. Repeat until vocabulary size reached\n\n**Used by**: ALBERT, T5, mBART, XLNet (via SentencePiece)\n\n```python\nfrom tokenizers import Tokenizer\nfrom tokenizers.models import Unigram\nfrom tokenizers.trainers import UnigramTrainer\n\ntokenizer = Tokenizer(Unigram())\n\ntrainer = UnigramTrainer(\n    vocab_size=8000,\n    special_tokens=[\"\u003Cunk>\", \"\u003Cs>\", \"\u003C\u002Fs>\"],\n    unk_token=\"\u003Cunk>\"\n)\n\ntokenizer.train(files=[\"data.txt\"], trainer=trainer)\n```\n\n**Advantages**:\n- Probabilistic (finds most likely tokenization)\n- Works well for languages without word boundaries\n- Handles diverse linguistic contexts\n\n**Trade-offs**:\n- Computationally expensive to train\n- More hyperparameters to tune\n\n## Tokenization pipeline\n\nComplete pipeline: **Normalization → Pre-tokenization → Model → Post-processing**\n\n### Normalization\n\nClean and standardize text:\n\n```python\nfrom tokenizers.normalizers import NFD, StripAccents, Lowercase, Sequence\n\ntokenizer.normalizer = Sequence([\n    NFD(),           # Unicode normalization (decompose)\n    Lowercase(),     # Convert to lowercase\n    StripAccents()   # Remove accents\n])\n\n# Input: \"Héllo WORLD\"\n# After normalization: \"hello world\"\n```\n\n**Common normalizers**:\n- `NFD`, `NFC`, `NFKD`, `NFKC` - Unicode normalization forms\n- `Lowercase()` - Convert to lowercase\n- `StripAccents()` - Remove accents (é → e)\n- `Strip()` - Remove whitespace\n- `Replace(pattern, content)` - Regex replacement\n\n### Pre-tokenization\n\nSplit text into word-like units:\n\n```python\nfrom tokenizers.pre_tokenizers import Whitespace, Punctuation, Sequence, ByteLevel\n\n# Split on whitespace and punctuation\ntokenizer.pre_tokenizer = Sequence([\n    Whitespace(),\n    Punctuation()\n])\n\n# Input: \"Hello, world!\"\n# After pre-tokenization: [\"Hello\", \",\", \"world\", \"!\"]\n```\n\n**Common pre-tokenizers**:\n- `Whitespace()` - Split on spaces, tabs, newlines\n- `ByteLevel()` - GPT-2 style byte-level splitting\n- `Punctuation()` - Isolate punctuation\n- `Digits(individual_digits=True)` - Split digits individually\n- `Metaspace()` - Replace spaces with ▁ (SentencePiece style)\n\n### Post-processing\n\nAdd special tokens for model input:\n\n```python\nfrom tokenizers.processors import TemplateProcessing\n\n# BERT-style: [CLS] sentence [SEP]\ntokenizer.post_processor = TemplateProcessing(\n    single=\"[CLS] $A [SEP]\",\n    pair=\"[CLS] $A [SEP] $B [SEP]\",\n    special_tokens=[\n        (\"[CLS]\", 1),\n        (\"[SEP]\", 2),\n    ],\n)\n```\n\n**Common patterns**:\n```python\n# GPT-2: sentence \u003C|endoftext|>\nTemplateProcessing(\n    single=\"$A \u003C|endoftext|>\",\n    special_tokens=[(\"\u003C|endoftext|>\", 50256)]\n)\n\n# RoBERTa: \u003Cs> sentence \u003C\u002Fs>\nTemplateProcessing(\n    single=\"\u003Cs> $A \u003C\u002Fs>\",\n    pair=\"\u003Cs> $A \u003C\u002Fs> \u003C\u002Fs> $B \u003C\u002Fs>\",\n    special_tokens=[(\"\u003Cs>\", 0), (\"\u003C\u002Fs>\", 2)]\n)\n```\n\n## Alignment tracking\n\nTrack token positions in original text:\n\n```python\noutput = tokenizer.encode(\"Hello, world!\")\n\n# Get token offsets\nfor token, offset in zip(output.tokens, output.offsets):\n    start, end = offset\n    print(f\"{token:10} → [{start:2}, {end:2}): {text[start:end]!r}\")\n\n# Output:\n# hello      → [ 0,  5): 'Hello'\n# ,          → [ 5,  6): ','\n# world      → [ 7, 12): 'world'\n# !          → [12, 13): '!'\n```\n\n**Use cases**:\n- Named entity recognition (map predictions back to text)\n- Question answering (extract answer spans)\n- Token classification (align labels to original positions)\n\n## Integration with transformers\n\n### Load with AutoTokenizer\n\n```python\nfrom transformers import AutoTokenizer\n\n# AutoTokenizer automatically uses fast tokenizers\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\n\n# Check if using fast tokenizer\nprint(tokenizer.is_fast)  # True\n\n# Access underlying tokenizers.Tokenizer\nfast_tokenizer = tokenizer.backend_tokenizer\nprint(type(fast_tokenizer))  # \u003Cclass 'tokenizers.Tokenizer'>\n```\n\n### Convert custom tokenizer to transformers\n\n```python\nfrom tokenizers import Tokenizer\nfrom transformers import PreTrainedTokenizerFast\n\n# Train custom tokenizer\ntokenizer = Tokenizer(BPE())\n# ... train tokenizer ...\ntokenizer.save(\"my-tokenizer.json\")\n\n# Wrap for transformers\ntransformers_tokenizer = PreTrainedTokenizerFast(\n    tokenizer_file=\"my-tokenizer.json\",\n    unk_token=\"[UNK]\",\n    pad_token=\"[PAD]\",\n    cls_token=\"[CLS]\",\n    sep_token=\"[SEP]\",\n    mask_token=\"[MASK]\"\n)\n\n# Use like any transformers tokenizer\noutputs = transformers_tokenizer(\n    \"Hello world\",\n    padding=True,\n    truncation=True,\n    max_length=512,\n    return_tensors=\"pt\"\n)\n```\n\n## Common patterns\n\n### Train from iterator (large datasets)\n\n```python\nfrom datasets import load_dataset\n\n# Load dataset\ndataset = load_dataset(\"wikitext\", \"wikitext-103-raw-v1\", split=\"train\")\n\n# Create batch iterator\ndef batch_iterator(batch_size=1000):\n    for i in range(0, len(dataset), batch_size):\n        yield dataset[i:i + batch_size][\"text\"]\n\n# Train tokenizer\ntokenizer.train_from_iterator(\n    batch_iterator(),\n    trainer=trainer,\n    length=len(dataset)  # For progress bar\n)\n```\n\n**Performance**: Processes 1GB in ~10-20 minutes\n\n### Enable truncation and padding\n\n```python\n# Enable truncation\ntokenizer.enable_truncation(max_length=512)\n\n# Enable padding\ntokenizer.enable_padding(\n    pad_id=tokenizer.token_to_id(\"[PAD]\"),\n    pad_token=\"[PAD]\",\n    length=512  # Fixed length, or None for batch max\n)\n\n# Encode with both\noutput = tokenizer.encode(\"This is a long sentence that will be truncated...\")\nprint(len(output.ids))  # 512\n```\n\n### Multi-processing\n\n```python\nfrom tokenizers import Tokenizer\nfrom multiprocessing import Pool\n\n# Load tokenizer\ntokenizer = Tokenizer.from_file(\"tokenizer.json\")\n\ndef encode_batch(texts):\n    return tokenizer.encode_batch(texts)\n\n# Process large corpus in parallel\nwith Pool(8) as pool:\n    # Split corpus into chunks\n    chunk_size = 1000\n    chunks = [corpus[i:i+chunk_size] for i in range(0, len(corpus), chunk_size)]\n\n    # Encode in parallel\n    results = pool.map(encode_batch, chunks)\n```\n\n**Speedup**: 5-8× with 8 cores\n\n## Performance benchmarks\n\n### Training speed\n\n| Corpus Size | BPE (30k vocab) | WordPiece (30k) | Unigram (8k) |\n|-------------|-----------------|-----------------|--------------|\n| 10 MB       | 15 sec          | 18 sec          | 25 sec       |\n| 100 MB      | 1.5 min         | 2 min           | 4 min        |\n| 1 GB        | 15 min          | 20 min          | 40 min       |\n\n**Hardware**: 16-core CPU, tested on English Wikipedia\n\n### Tokenization speed\n\n| Implementation | 1 GB corpus | Throughput    |\n|----------------|-------------|---------------|\n| Pure Python    | ~20 minutes | ~50 MB\u002Fmin    |\n| HF Tokenizers  | ~15 seconds | ~4 GB\u002Fmin     |\n| **Speedup**    | **80×**     | **80×**       |\n\n**Test**: English text, average sentence length 20 words\n\n### Memory usage\n\n| Task                    | Memory  |\n|-------------------------|---------|\n| Load tokenizer          | ~10 MB  |\n| Train BPE (30k vocab)   | ~200 MB |\n| Encode 1M sentences     | ~500 MB |\n\n## Supported models\n\nPre-trained tokenizers available via `from_pretrained()`:\n\n**BERT family**:\n- `bert-base-uncased`, `bert-large-cased`\n- `distilbert-base-uncased`\n- `roberta-base`, `roberta-large`\n\n**GPT family**:\n- `gpt2`, `gpt2-medium`, `gpt2-large`\n- `distilgpt2`\n\n**T5 family**:\n- `t5-small`, `t5-base`, `t5-large`\n- `google\u002Fflan-t5-xxl`\n\n**Other**:\n- `facebook\u002Fbart-base`, `facebook\u002Fmbart-large-cc25`\n- `albert-base-v2`, `albert-xlarge-v2`\n- `xlm-roberta-base`, `xlm-roberta-large`\n\nBrowse all: https:\u002F\u002Fhuggingface.co\u002Fmodels?library=tokenizers\n\n## References\n\n- **[Training Guide](references\u002Ftraining.md)** - Train custom tokenizers, configure trainers, handle large datasets\n- **[Algorithms Deep Dive](references\u002Falgorithms.md)** - BPE, WordPiece, Unigram explained in detail\n- **[Pipeline Components](references\u002Fpipeline.md)** - Normalizers, pre-tokenizers, post-processors, decoders\n- **[Transformers Integration](references\u002Fintegration.md)** - AutoTokenizer, PreTrainedTokenizerFast, special tokens\n\n## Resources\n\n- **Docs**: https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftokenizers\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftokenizers ⭐ 9,000+\n- **Version**: 0.20.0+\n- **Course**: https:\u002F\u002Fhuggingface.co\u002Flearn\u002Fnlp-course\u002Fchapter6\u002F1\n- **Paper**: BPE (Sennrich et al., 2016), WordPiece (Schuster & Nakajima, 2012)\n",{"name":5,"author":44,"license":15,"version":45,"metadata":46,"platforms":50,"description":7,"dependencies":51},"Orchestra Research","1.0.0",{"hermes":47},{"tags":48},[28,19,27,26,25,24,49,23,22,31],"Rust",[38,30,35],[52,53,29],"tokenizers","transformers","main","0d1434f89f4bc6b07e9d272accd8c1013768820f","archive\u002Fskill-md-reviewed\u002Fsha256\u002F79\u002F79ca48477b7b72a9d5b59e472c89bec4399048ba2f5e9cc3698ed6e6facc5b0c",0,"https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills","https:\u002F\u002Fgithub.com\u002Fzapabob","The Open University of JAPAN　2019-now\r\n\r\nTokyo University of Science　2007-2010","2026-10-26 16:29:01",{"source":8,"skillName":5,"commands":63,"prompt":73,"zipAvailable":74},[64,67,70],{"tool":65,"command":66},"npx","npx skills add zapabob\u002FSkills --skill huggingface-tokenizers",{"tool":68,"command":69},"bunx","bunx skills add zapabob\u002FSkills --skill huggingface-tokenizers",{"tool":71,"command":72},"pnpm","pnpm dlx skills add zapabob\u002FSkills --skill huggingface-tokenizers","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Farchive\u002Fskill-md-reviewed\u002Fsha256\u002F79\u002F79ca48477b7b72a9d5b59e472c89bec4399048ba2f5e9cc3698ed6e6facc5b0c\u002FSKILL.md",true,[],[77,112,147,173,200,217],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":96,"publishedAt":39},"01M6GJTE1GESGGF9K028Y7753E","xurl","X\u002FTwitter via xurl CLI: post, search, DM, media, v2 API.","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Fregistry\u002Fskills\u002Fxurl\u002Fvariants\u002Fhermes-agent\u002FSKILL.md",[83,88,91],{"categoryNo":84,"categoryName":85,"taxonomyType":86,"taxonomyTypeDesc":87},"01JBSEEDCAT000000000000010","后端与接口","SCENARIO","使用场景",{"categoryNo":89,"categoryName":90,"taxonomyType":86,"taxonomyTypeDesc":87},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":92,"categoryName":93,"taxonomyType":94,"taxonomyTypeDesc":95},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[79,97,30,98,99,38,35,100,101,102,103,104,105,106,107,108,54,109,110,111],"openclaw","cli","github","official","search","https","social-media","api","Media","hermes","hermes agent","twitter","social","via","post",{"skillNo":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":116,"categories":117,"keywords":129,"publishedAt":39},"01M6GJT8S1PXQ8WPXT1HSNBWH2","writing-plans","Write implementation plans: bite-sized tasks, paths, code.","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Fregistry\u002Fskills\u002Fwriting-plans\u002Fvariants\u002Fhermes-agent\u002FSKILL.md",[118,121,124,125,126],{"categoryNo":119,"categoryName":120,"taxonomyType":86,"taxonomyTypeDesc":87},"01JBSEEDCAT000000000000015","视觉设计",{"categoryNo":122,"categoryName":123,"taxonomyType":86,"taxonomyTypeDesc":87},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":89,"categoryName":90,"taxonomyType":86,"taxonomyTypeDesc":87},{"categoryNo":92,"categoryName":93,"taxonomyType":94,"taxonomyTypeDesc":95},{"categoryNo":127,"categoryName":128,"taxonomyType":94,"taxonomyTypeDesc":95},"01JBSEEDCAT000000000000036","内容创作者",[130,131,132,133,30,134,38,135,35,136,137,138,139,140,141,142,143,144,145,146],"writing","writing plans","plans","design","review","workflow","documentation","tasks","development","code","planning","code-review","subagent","write","superpowers","paths","test-driven-development",{"skillNo":148,"name":149,"nameZh":6,"description":150,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":151,"categories":152,"keywords":156,"publishedAt":39},"01M6GJST0H5X7WHBJ3BZK0DVKH","whisper","OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual ASR.","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Fregistry\u002Fskills\u002Fwhisper\u002Fvariants\u002Fhermes-agent-optional\u002FSKILL.md",[153],{"categoryNo":154,"categoryName":155,"taxonomyType":86,"taxonomyTypeDesc":87},"01JBSEEDCAT000000000000016","图像与多媒体",[149,30,38,157,158,159,32,160,161,162,163,164,165,166,167,168,169,170,171,172],"openai","speech","multimodal","ASR","Audio","text","model","speech recognition","use","general","multilingual","translation","english","identification","processing","Audio Processing",{"skillNo":174,"name":175,"nameZh":6,"description":176,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":177,"categories":178,"keywords":182,"publishedAt":39},"01M6GJSMRQ8EETZTYMG72P70EA","weights-and-biases","W&B: log ML experiments, sweeps, model registry, dashboards.","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Fregistry\u002Fskills\u002Fweights-and-biases\u002Fvariants\u002Fhermes-agent\u002FSKILL.md",[179],{"categoryNo":180,"categoryName":181,"taxonomyType":86,"taxonomyTypeDesc":87},"01JBSEEDCAT000000000000013","分析与可视化",[183,30,184,185,35,186,187,188,189,190,191,163,192,193,194,195,196,197,198,199],"weights and biases","tensorflow","experiments","pytorch","visualization","registry","real-time","tuning","dashboards","experiment","log","time","Experiment Tracking","Model Registry","Hyperparameter Tuning","WandB","Real-Time Visualization",{"skillNo":201,"name":202,"nameZh":6,"description":203,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":204,"categories":205,"keywords":207,"publishedAt":39},"01M6GJS19KDPRY112JTR326S06","watchers","Poll RSS, JSON APIs, and GitHub with watermark dedup.","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Fregistry\u002Fskills\u002Fwatchers\u002Fvariants\u002Fhermes-agent-optional\u002FSKILL.md",[206],{"categoryNo":89,"categoryName":90,"taxonomyType":86,"taxonomyTypeDesc":87},[202,30,208,99,38,209,210,211,212,213,214,106,107,215,216],"terminal","automation","http","monitoring","cron","json","rss","devops","with",{"skillNo":218,"name":219,"nameZh":6,"description":220,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":221,"categories":222,"keywords":225,"publishedAt":39},"01M6GJRHJD2A6Q830SWF9W48BC","serving-llms-vllm","vLLM: high-throughput LLM serving, OpenAI API, quantization.","https:\u002F\u002Fgithub.com\u002Fzapabob\u002FSkills\u002Fblob\u002Fmain\u002Fregistry\u002Fskills\u002Fvllm\u002Fvariants\u002Fhermes-agent\u002FSKILL.md",[223,224],{"categoryNo":84,"categoryName":85,"taxonomyType":86,"taxonomyTypeDesc":87},{"categoryNo":92,"categoryName":93,"taxonomyType":94,"taxonomyTypeDesc":95},[226,227,228,38,36,157,229,230,231,32,31,104,232,233,234,235,236,237,238,239],"llms","vLLM","serving llms vllm","openai api","tensor","Inference Serving","llm serving","High Throughput","PagedAttention","Continuous Batching","Quantization","Tensor Parallelism","high-throughput","throughput"]