The tokenization pipeline

When calling encode() or encode_batch(), the input text(s) go through the following pipeline:

We’ll see in details what happens during each of those steps in detail, as well as when you want to decode some token ids, and how the 🤗 Tokenizers library allows you to customize each of those steps to your needs. If you’re already familiar with those steps and want to learn by seeing some code, jump to our BERT from scratch example.

For the examples that require a Tokenizer, we will use the tokenizer we trained in the Quicktour, which you can load with:

    from tokenizers import Tokenizer

    tokenizer = Tokenizer.from_file("data/tokenizer-wiki.json")


Normalization is, in a nutshell, a set of operations you apply to a raw string to make it less random or “cleaner”. Common operations include stripping whitespace, removing accented characters or lowercasing all text. If you’re familiar with Unicode normalization, it is also a very common normalization operation applied in most tokenizers.

Each normalization operation is represented in the 🤗 Tokenizers library by a Normalizer, and you can combine several of those by using a Sequence. Here is a normalizer applying NFD Unicode normalization and removing accents as an example:

from tokenizers import normalizers
from tokenizers.normalizers import NFD, StripAccents

normalizer = normalizers.Sequence([NFD(), StripAccents()])

You can manually test that normalizer by applying it to any string:

normalizer.normalize_str("Héllò hôw are ü?")
# "Hello how are u?"

When building a Tokenizer, you can customize its normalizer by just changing the corresponding attribute:

tokenizer.normalizer = normalizer

Of course, if you change the way a tokenizer applies normalization, you should probably retrain it from scratch afterward.


Pre-tokenization is the act of splitting a text into smaller objects that give an upper bound to what your tokens will be at the end of training. A good way to think of this is that the pre-tokenizer will split your text into “words” and then, your final tokens will be parts of those words.

An easy way to pre-tokenize inputs is to split on spaces and punctuations, which is done by the Whitespace pre-tokenizer:

from tokenizers.pre_tokenizers import Whitespace

pre_tokenizer = Whitespace()
pre_tokenizer.pre_tokenize_str("Hello! How are you? I'm fine, thank you.")
# [("Hello", (0, 5)), ("!", (5, 6)), ("How", (7, 10)), ("are", (11, 14)), ("you", (15, 18)),
#  ("?", (18, 19)), ("I", (20, 21)), ("'", (21, 22)), ('m', (22, 23)), ("fine", (24, 28)),
#  (",", (28, 29)), ("thank", (30, 35)), ("you", (36, 39)), (".", (39, 40))]

The output is a list of tuples, with each tuple containing one word and its span in the original sentence (which is used to determine the final offsets of our Encoding). Note that splitting on punctuation will split contractions like "I'm" in this example.

You can combine together any PreTokenizer together. For instance, here is a pre-tokenizer that will split on space, punctuation and digits, separating numbers in their individual digits:

from tokenizers import pre_tokenizers
from tokenizers.pre_tokenizers import Digits

pre_tokenizer = pre_tokenizers.Sequence([Whitespace(), Digits(individual_digits=True)])
pre_tokenizer.pre_tokenize_str("Call 911!")
# [("Call", (0, 4)), ("9", (5, 6)), ("1", (6, 7)), ("1", (7, 8)), ("!", (8, 9))]

As we saw in the Quicktour, you can customize the pre-tokenizer of a Tokenizer by just changing the corresponding attribute:

tokenizer.pre_tokenizer = pre_tokenizer

Of course, if you change the way the pre-tokenizer, you should probably retrain your tokenizer from scratch afterward.

The Model

Once the input texts are normalized and pre-tokenized, the Tokenizer applies the model on the pre-tokens. This is the part of the pipeline that needs training on your corpus (or that has been trained if you are using a pretrained tokenizer).

The role of the model is to split your “words” into tokens, using the rules it has learned. It’s also responsible for mapping those tokens to their corresponding IDs in the vocabulary of the model.

This model is passed along when intializing the Tokenizer so you already know how to customize this part. Currently, the 🤗 Tokenizers library supports:

For more details about each model and its behavior, you can check here


Post-processing is the last step of the tokenization pipeline, to perform any additional transformation to the Encoding before it’s returned, like adding potential special tokens.

As we saw in the quick tour, we can customize the post processor of a Tokenizer by setting the corresponding attribute. For instance, here is how we can post-process to make the inputs suitable for the BERT model:

from tokenizers.processors import TemplateProcessing

tokenizer.post_processor = TemplateProcessing(
    single="[CLS] $A [SEP]",
    pair="[CLS] $A [SEP] $B:1 [SEP]:1",
    special_tokens=[("[CLS]", 1), ("[SEP]", 2)],

Note that contrarily to the pre-tokenizer or the normalizer, you don’t need to retrain a tokenizer after changing its post-processor.

All together: a BERT tokenizer from scratch

Let’s put all those pieces together to build a BERT tokenizer. First, BERT relies on WordPiece, so we instantiate a new Tokenizer with this model:

from tokenizers import Tokenizer
from tokenizers.models import WordPiece

bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))

Then we know that BERT preprocesses texts by removing accents and lowercasing. We also use a unicode normalizer:

from tokenizers import normalizers
from tokenizers.normalizers import Lowercase, NFD, StripAccents

bert_tokenizer.normalizer = normalizers.Sequence([NFD(), Lowercase(), StripAccents()])

The pre-tokenizer is just splitting on whitespace and punctuation:

from tokenizers.pre_tokenizers import Whitespace

bert_tokenizer.pre_tokenizer = Whitespace()

And the post-processing uses the template we saw in the previous section:

from tokenizers.processors import TemplateProcessing

bert_tokenizer.post_processor = TemplateProcessing(
    single="[CLS] $A [SEP]",
    pair="[CLS] $A [SEP] $B:1 [SEP]:1",
        ("[CLS]", 1),
        ("[SEP]", 2),

We can use this tokenizer and train on it on wikitext like in the Quicktour:

from tokenizers.trainers import WordPieceTrainer

trainer = WordPieceTrainer(
    vocab_size=30522, special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
files = [f"data/wikitext-103-raw/wiki.{split}.raw" for split in ["test", "train", "valid"]]
bert_tokenizer.train(files, trainer)"data/bert-wiki.json")


On top of encoding the input texts, a Tokenizer also has an API for decoding, that is converting IDs generated by your model back to a text. This is done by the methods decode() (for one predicted text) and decode_batch() (for a batch of predictions).

The decoder will first convert the IDs back to tokens (using the tokenizer’s vocabulary) and remove all special tokens, then join those tokens with spaces:

output = tokenizer.encode("Hello, y'all! How are you 😁 ?")
# [1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2]

tokenizer.decode([1, 27253, 16, 93, 11, 5097, 5, 7961, 5112, 6218, 0, 35, 2])
# "Hello , y ' all ! How are you ?"

If you used a model that added special characters to represent subtokens of a given “word” (like the "##" in WordPiece) you will need to customize the decoder to treat them properly. If we take our previous bert_tokenizer for instance the default decoing will give:

output = bert_tokenizer.encode("Welcome to the 🤗 Tokenizers library.")
# ["[CLS]", "welcome", "to", "the", "[UNK]", "tok", "##eni", "##zer", "##s", "library", ".", "[SEP]"]

# "welcome to the tok ##eni ##zer ##s library ."

But by changing it to a proper decoder, we get:

from tokenizers import decoders

bert_tokenizer.decoder = decoders.WordPiece()
# "welcome to the tokenizers library."