TkTkT: the ToKeniser ToolKiT

A collection of Pythonic subword tokenisers and text preprocessing tools, with full backwards- and forwards-compatibility with HuggingFace tokenizers. One package to rule them all.

Quick navigation:

• Installation
• Features
• Examples
• Architecture

Features

Supported tokenisers

All subword tokenisers are defined under tktkt.models. Many of these can be instantiated without much background knowledge using the factory classes in tktkt.factories. Also, any HuggingFace tokeniser can be wrapped into a TkTkT tokeniser, and any TkTkT tokeniser can be wrapped into a HuggingFace tokeniser.

Currently, the package implements:

• Byte-pair encoding (BPE) tokenisers:
  • Classical BPE (Sennrich et al., 2016), with added support for any word boundary marker (Ġ, _, </w>, ...) and n-ary merges (byte-tuple encoding, BTE).
  • BPE-dropout (Provilkov et al., 2020)
  • BPE-knockout (Bauwens & Delobelle, 2024)
  • PickyBPE (Chizhov et al., 2024)
  • ScaffoldBPE (Lian et al., 2025)
  • TrimmedBPE (Cognetta et al., 2024)
  • Other experimental variants I implemented just for fun:
    • BPE-breakdown: BPE which starts randomly undoing merges after it finishes deterministically, similar to StochasTok.
    • Non-geometric BPE-dropout: BPE-dropout, but rather than picking merges geometrically, picks them uniformly.
    • EnsuredBPE: BPE where the last merges have been replaced by the merges necessary to ensure that a given list of strings is in the vocabulary.
    • ShuffledBPE: BPE but with merge priorities shuffled, although types are never shuffled to a priority before the ancestors in their merge tree.
• Unigram language model (ULM), dubbed KudoPiece in TkTkT (Kudo, 2018):
  • Wrapper around the SentencePiece package, or
  • Native implementation in TkTkT
• Greedy tokenisers:
  • MaxMatch (Hiraoka, 2022), a.k.a. left-to-right greedy tokenisation, and also right-to-left (Bauwens, 2023 and later Uzan et al., 2024)
  • FLOTA (Hofmann et al., 2022), i.e. random-access longest-first tokenisation.
  • Other experimental variants:
    • Last-BPE-first: random-access youngest-first tokenisation (specifically for BPE vocabularies).
    • Left-to-right-to-left greedy: L2R2L_Greedy
• Subword regularisers:
  • GRaMPa (Bauwens et al., 2025): randomised segmentation constrained by a vocabulary.
  • StochasTok (Sims et al., 2025)
• SaGe (Yehezkel & Pinter, 2023) vocabularisation.
• Derivative leverager (DeL) (Hofmann et al., 2021), both training and segmentation.
• Other, less interesting tokenisers:
  • Character/byte N-grams.
  • Lempel-Ziv-Welch (LZW) as a tokeniser (Zouhar et al., 2023).

Currently work in progress:

• Less-is-Better (LiB)
• Tokenmonster
• DPE
• Morfessor family
• VOLT (mostly already present in TkTkT)

Multiplexing

TkTkT is the only package that supports multiplexing multiple tokenisers into one big tokeniser that alternates between each of them. There are multiplexers that do this deterministically (e.g. choosing the tokeniser that compresses the input the most) or stochastically (e.g. choosing among a set of tokenisers uniformly).

Evaluation metrics

TkTkT's evaluation framework aims to do as little work as possible. It can dispatch tokens produced by a tokeniser to as many metrics as you need at once, and caches everything it can so you don't have to compute any metric twice. See here for an example.

TkTkT currently supports the following intrinsic tokeniser evaluation metrics:

• Fertility statistics: how many tokens the tokeniser produces per word, and how many segmentations its vocabulary could produce in theory.
• Morphological boundary recognition: using the tokeniser as a binary classifier for whether two morphemes meet at each position in a word.
• Information-theoretic measures, including Rényi entropy and Rényi efficiency.
• Window-based metrics like MATTR.
• Bigram metrics to quantify the richness of token contexts, like accessor variety.
• Comparisons between two tokenisers: how much they tokenise words exactly the same, and how much their split points overlap.

Security

It should be impossible for users to jailbreak a language model by forcing its tokeniser to produce a special token (e.g. <|endoftext|>, system prompt delimiters, ...). The reason other packages do not have this guarantee is that they represent special tokens as strings and then give the tokeniser access to these strings. In TkTkT, subword vocabularies are objects that hide special tokens from their tokeniser. In fact, special tokens are defined as integers, not as strings.

When loading a tokeniser trained in another package that probably inserted specials into the vocabulary, TkTkT explicitly requires the user to declare which strings are actually specials, and hides them from the tokeniser.

User-friendliness

Caching

When a tokeniser finishes training, you shouldn't be forced to mess with file paths to connect your training and testing scripts together. In TkTkT, training code caches its results. When you rerun it, it will load its results from disk and skip the waiting time.

Type-checking

I really f*cking passionately hate when my IDE cannot perform autocompletion because of poor type annotation or design. People have been suffering under the idiocy of AutoTokenizer for too long. In TkTkT, everything is as type-annotated as possible, which means:

• There are no checkpoint strings in TkTkT. There are Artifacts objects, which not only declare how to get the results of training, but also which Preprocessor was used, so that complex preprocessing objects are just known and never have to be stringified.
• When loading a tokeniser using a TokeniserFactory and Artifacts (the equivalent of AutoTokenizer of a str), the exact type of tokeniser is known.
• The special tokens in a vocabulary appear in autocompletion. For example, for BERT's BPE tokeniser, tokeniser.vocab.specials. will show CLS, SEP, PAD, MASK in your IDE.
• After training a tokeniser, the results on disk are already parsed into an object for you. When you train a BPE tokeniser, the results that come out aren't a dumb file path. It's an Artifacts object that has a .getVocabulary() and .getMerges() method.

Preprocessing

TkTkT has a rich set of text mappings and pretokenisers that preprocess text before it is tokenised, including support for stochastic perturbation. Unlike other libraries, preprocessors are objects, not regular expressions. This allows much more powerful processing than regex, whilst being more easy to read. See if you can understand this arguably complicated transformation:

from tktkt.preparation.splitters import *
from tktkt.preparation.mappers import PseudoByteMapping
from tktkt.factories.preprocessors import RobertaSpaceMarker


class ExamplePretokeniser(PretokeniserSequence):
    def __init__(self):
        super().__init__([
            IsolatePunctuation(HyphenMode.EXCLUDED, protect_apostrophes_without_spaces=True),
            OnWhitespace(destructive=True),
            IsolateEnglishContractions(do_nt=True),

            MapperAsPretokeniser(PseudoByteMapping()),
            AddWordBoundary(RobertaSpaceMarker),

            IsolateDigits(),
            IsolatePunctuation(HyphenMode.ONLY)
        ])

TkTkT also comes with language-specific pretokenisation like Japanese word segmentation and Thai word segmentation.

Visualisers

The following tokenisation procedures can be visualised:

• BPE/BTE: the final merge tree (in regular LaTeX), as well as an animated progression of the merges (in LaTeX Beamer).

Opinionated

Apart from the type-checking and caching described above, TkTkT enforces several truths about tokenisation which are not present in other packages:

• Tokenisers produce string segments (tokens), not integer identifiers (IDs). There is no "BPE way" of mapping ["_un", "believ", "able"] to integers. That should be done by a separate object, the Vocab.
• Preprocessors should not (only) be regular expressions. They should be chains of Python code.
• Spaces should not be treated as word boundaries. If a word does not have a "prefix space", e.g. because it is the start of the sentence or because it is preceded by a punctuation mark, it should still receive a boundary character, and the developer should be able to decide if that boundary comes at the start or the end of the word.
• Special tokens (CLS, SEP, BOS, ...) should not have a string representation ("[CLS]", "[SEP]", "<s>", ...), or at least they should not be in the vocabulary, or at least they should not be used by the tokeniser. All of these are violated by HuggingFace transformers (see the bottom of this README).
• Preprocessing should not be baked into the tokeniser.
  • Tokenisers should segment the given string, not transform it into some unpredictably other string thing and segment that one.
  • Any parameters regarding preprocessing should not be in the tokeniser's constructor, except for the actual Preprocessor itself.
• Preprocessors should not be pickled and base64'ed into opaque objects. They should be defined as Python code.

Architecture

Main interfaces

The goal of TkTkT is to provide a straightforward Pythonic interface for everything-tokenisation, and to be as object-oriented as possible. The main interfaces are found under tktkt.interfaces.

Inference

Fundamentally, all tokenisers are a Tokeniser that have a Preprocessor.

• The Tokeniser class has two important methods:

  • .tokenise(pretoken: str) -> List[str]: segments a string as-is into parts.
  • .prepareAndTokenise(text: str) -> List[str]: applies the tokeniser's preprocessor and then tokenises each pre-token separately.

• The Preprocessor class is a pipeline of three components:

  1. a non-invertible text mapping
  2. an invertible text mapping
  3. a pretokeniser that splits strings into smaller strings.

To map tokens (string segments) to identifiers (integers) for indexing into an embedding matrix, this interface is extended to the TokeniserWithVocabulary. This class makes use of a Vocab object which keeps types used by the tokeniser separated from the specials used by a downstream language model. This prevents token injection attacks, which HuggingFace transformers/tokenizers is vulnerable to (see the bottom of this README).

Training

To learn the parameters of a Tokeniser (e.g. BPE merges), there is the Vocabulariser class. It can learn from word-count files or from corpora of sentences. It takes a Preprocessor exactly like Tokeniser, except Vocabulariser is for training the tokeniser (vocabularisation) and Tokeniser is for inference (segmentation).

Loading

To make it easier to load the results of a vocabularisation run from storage back into Python, there are Artifacts classes to do this for you.

For ease-of-use, many Tokeniser classes have a TokeniserFactory defined for them that simplify the instantiation process. Often, a TokeniserFactory will take an Artifacts object to provide it any files.

Submodules

The packages is divided into the following submodules:

• tktkt.interfaces: contains the main parent classes from which all other classes derive.
  • The most important classes are TextMapper, Pretokeniser, Preprocessor, Vocabulariser, Tokeniser, Artifacts, and TokeniserFactory.
• tktkt.preparation: contains all the text preprocessing tools.
• tktkt.models: contains all the tokenisation (i.e. vocabularisation and/or segmentation) algorithms.
• tktkt.evaluation: contains procedures with which to quantify a Tokeniser through inference.
• tktkt.factories: contains a bunch of pre-defined constructor calls, for both vocabularies and tokenisers:
  • tktkt.factories.artifacts: contains classes that load the files for specific tokenisers.
  • tktkt.factories.tokenisers: contains tokeniser factories.
  • tktkt.factories.specials: contains sets of special identifiers (CLS, SEP, BOS, EOS, ...).
  • tktkt.factories.preprocessors: contains a bunch of pre-defined preprocessors so you don't have to. Check out the ModernEnglishPreprocessor, for example.
  • tktkt.factories.evaluation: contains pre-built tokeniser evaluation pipelines.
• tktkt.wrappers: contains classes that wrap around existing tokenisers to equip them with more features.
  • tktkt.wrappers.multiplexing: alternate between multiple tokenisers within the same sentence.
  • tktkt.wrappers.hashingvocab: add a string-to-integer mapping to a Tokeniser that can produce any substring, turning it into a TokeniserWithFiniteIdRange.
• tktkt.visualisation: contains procedures to generate explanatory LaTeX code about some models.
• tktkt.util: contains tools peripheral to tokenisation, like string formatting, combinatoric calculations, iterable functions, timing, etc...

Installation

To just get TkTkT and all its dependencies, run

pip install tktkt[all]

If you also want to make use of the HuggingFace compatibility, you'll need transformers, which can be installed for you if you don't already have it by running

pip install tktkt[all,hf]

Examples

HuggingFace compatibility

In the example below, a BPE tokeniser is loaded from the HuggingFace hub as a PreTrainedTokenizerFast and converted into a TkTkT Tokeniser object. Then, this object is itself converted into a HuggingFace PreTrainedTokenizer again.

# Backwards-compatibility:
from transformers import AutoTokenizer
from tktkt.models.huggingface.wrapper import HuggingFaceTokeniser

hf_roberta = AutoTokenizer.from_pretrained("roberta-base")
tktkt_roberta = HuggingFaceTokeniser(hf_roberta)

###
sentence = " That's so supercalifragilisticexpialidocious, Günther!"
print("Full tokenisation pipeline:")
print("\tHF Tk:", hf_roberta.tokenize(sentence))  # Note the lack of autocompletion on this.
print("\tTkTkT:", tktkt_roberta.prepareAndTokenise(sentence))
print("Only the preprocessing:")
print("\tTkTkT:", tktkt_roberta.preprocessor.do(sentence))
###

# Forwards-compatibility:
from tktkt.interfaces.huggingface import TktktToHuggingFace

hf_tktkt_roberta = TktktToHuggingFace(tktkt_roberta, specials_from=hf_roberta)
print(hf_tktkt_roberta.tokenize(sentence))

Training and instantiating BPE

Here's a minimal working example to train a BPE tokeniser on the first 100 000 examples of an English Wikipedia dataset:

from datasets import load_dataset
from tktkt.factories.preprocessors import ModernEnglishPreprocessor, KudoSpaceMarker
from tktkt.models.bpe.vocabularisation import BPEVocabulariser

corpus        = load_dataset("Salesforce/wikitext", "wikitext-103-raw-v1", split="train", streaming=True).take(100_000)
preprocessor  = ModernEnglishPreprocessor(marker=KudoSpaceMarker)
vocabulariser = BPEVocabulariser(preprocessor=preprocessor, vocab_size=32_768)
results       = vocabulariser.vocabulariseFromHf(corpus, text_field="text")

That's just 7 lines of code to get a tokeniser from a corpus! To load the result into a HuggingFace-accelerated tokeniser, we can call

from tktkt.models.huggingface.bpe import HuggingFaceBPETokeniser

tokeniser = HuggingFaceBPETokeniser(
    preprocessor=results.preprocessorEffective(),
    vocab=results.getVocabulary(),
    merges=results.getMerges()
)

The best part? You can put this in the same script as the vocabulariser.vocabulariseFromHf, right below it. All training methods are cached, which means that if the result was computed before, it will be loaded from disk immediately.

If you want to load the result of training in a different project, the way to do it is to first write a BPEArtifacts class that knows which preprocessor you used and where to find your results. Once you have this, use the Factory_BPE to read the artifacts for you.

As an example: I once trained BPE across the first 3 million examples in SlimPajama. The BPEArtifacts class for this specific training run are predefined in TkTkT:

from tktkt.factories.artifacts import BPE32ki_SlimPajama3M
from tktkt.factories.tokenisers import Factory_BPE

tokeniser = Factory_BPE(files=BPE32ki_SlimPajama3M()).buildTokeniser()

Note again that the preprocessor comes with the artifacts, so the factory doesn't require that you specify it.

Training and instantiating ULM (a.k.a. KudoPiece)

Let's now say you want to train and load an English ULM tokeniser. You are, of course, scared of the sentencepiece library because its Python interface is a thin wrapper around a command-line call, not allowing autocompletion in your IDE. In TkTkT, you would proceed as follows (note that ULM is called "KudoPiece" in TkTkT because many tokenisers are based on a language model of unigrams).

First we instantiate a preprocessor, and call the trainer with relevant training arguments. Then, we load the saved results into a tokeniser.

from tktkt.factories.preprocessors import ModernEnglishPreprocessor_SentencePieceCompatible, BoundaryMarkerLocation
from tktkt.models.kudopiece.vocabularisation import *

### Your data iterator goes here.
sentence_corpus: Iterable[str] = ...
###

preprocessor = ModernEnglishPreprocessor_SentencePieceCompatible(
    marker_location=BoundaryMarkerLocation.START
)

trainer = KudoPieceVocabulariser(
    preprocessor=preprocessor,
    final_vocab_size=40_000,
    arguments=KudoPieceArguments(character_coverage=0.9995),
    file_stem="tutorial"
)
results = trainer.vocabulariseFromHf(sentence_corpus, text_field="text")


from tktkt.models.kudopiece.segmentation import KudoPieceTokeniser

tokeniser = KudoPieceTokeniser(preprocessor=results.preprocessorNative(), model_file=results.getModelFile())
print(tokeniser.prepareAndTokenise("Hello there, my good friend!"))

Custom preprocessing

TkTkT preprocesses text into pretokens not with a regular expression, but with a sequence of Python objects that can perform any operation they want on the current pretokens. It is hence strictly more expressive than regex-based pretokenisation. For example:

from tktkt.factories.preprocessors import *

toy_preprocessor = Preprocessor(
    Lowercaser(),
    Replace("!", "."),
    PretokeniserSequence([
        OnWhitespace(),
        IsolatePunctuation(),
        AddWordBoundary(KudoSpaceMarker)
    ])
)

print(toy_preprocessor.do("This example will be preprocessed (even without a tokeniser)!"))

This can then be used to instantiate any TkTkT tokeniser, whose functionality is decoupled from the preprocessor. For example:

from tktkt.models.greedy.directional import L2R_Greedy, Vocab
from tktkt.factories.specials import NoSpecials

tokeniser = L2R_Greedy(
    preprocessor=toy_preprocessor,
    vocab=Vocab(
        ["a", "b", "c", "d", "ab", "ba", ".", ",", "▁"],
        specials=NoSpecials(),
        unk_id=0
    )
)

print(tokeniser.prepareAndTokenise("A bad cab, ABBA!"))
print(tokeniser.tokenise("abc."))

There are many more preprocessing classes available, some pre-made. Check out the ModernEnglishPreprocessor for typical modern use-cases.

Why does this package exist if we have HuggingFace tokenizers?

First of all, note again that TkTkT has backwards compatibility with HuggingFace tokenizers. There are wrapper classes for tokenisers under tktkt.models.huggingface and for normalisers/pretokenisers under tktkt.preparation.huggingface.

Note also that the tokenizers package actually lacks most of the logic which the HuggingFace ecosystem uses for tokenisation. Both the main tokenisation interface (PretrainedTokenizer/PretrainedFastTokenizer) as well as deserialisation (AutoTokenizer) live in transformers, not tokenizers. So let's consider the union of both.

Here's a non-exhaustive list of reasons:

1. The tokenizers package offers very few core models (basically only BPE and KudoPiece, which sentencepiece already offers and actually keeps updated), yet there exist many more in the literature. The likelihood that someone who knows the literature comes along to implement all of them in Rust or C++ is rather low.
2. The HuggingFace tokenizers package has horrifically un(der)documented Python interfaces.
   • Some tokenizers classes even accept arguments that aren't in their signature.
   • Your IDE cannot know what type the result of AutoTokenizer is, and thus you can never have autocompletion on it.
3. The tokenizers package is implemented in Rust. This has at least two big issues with it:
   • There is no possibility of inspecting implementations in any Python IDE.
   • Some of the responsibilities of existing Python code in transformers for tokenisation are unfortunately delegated to the Rust codebase when it is used, and thus re-implemented in tokenizers. This means HuggingFace should maintain parallel implementations, but of course, Python and Rust engineers are different people, and thus the codebases have become desynchronised. So, some constructor arguments do work for some tokenisers whilst doing nothing for others.
   • One example of a bug: the tokenizers package's RoBERTa BPE implementation removes the highest-priority merge from the tokeniser unless the merge file is preceded by a #version tag? This doesn't conform to the BPE standard, and almost cost me a paper.
4. The transformers package forces special tokens (<|endoftext|>, [CLS], [SEP], ...) to be treated as if they are user input. That's a security vulnerability.
   • Special tokens should never be treated like text. They should be seen as IDs without a name. They are purely for adding extra embedding vectors to the input of a downstream language model.
   • Yet, in the PreTrainedTokenizerBase class, specials must be declared using only a string with no identifier, and the point at which these strings receive their identifier is when they are run through the exact same method that converts the tokens from user input to identifiers.
5. The tokenizers interface does not allow separating preprocessing from the actual tokenisation algorithm.
   • The PreTrainedTokenizerBase class, from which the "slow" (Pythonic) PreTrainedTokenizer and "fast" (Rustic) PreTrainedTokenizerFast classes both inherit, only declares an end-to-end .tokenize() method (equivalent to TkTkT's .prepareAndTokenise()). The interface for these subclasses is different enough that both lack features of the other:
     • Whereas PreTrainedTokenizer does declare a ._tokenize() (equivalent to TkTkT's .tokenise()), I challenge you to find the equivalent for PreTrainedTokenizerFast. Best you'll find is .backend_tokenizer.model.tokenize(), which outputs unusable objects of class tokenizers.Token.
     • Whereas PreTrainedTokenizerFast has fields .backend_tokenizer.pre_tokenizer and .backend_tokenizer.normalizer (untyped of course, so you can't get autocompletion on their methods unless you manually assign them to a variable and annotate it yourself), PreTrainedTokenizer has no access to a pretokeniser. Preprocessing has to be defined inside ._tokenize(), which means you're doing two steps of preprocessing (one inside .tokenize() and one inside ._tokenize()) making this ._tokenize() no longer equivalent to TkTkT's .tokenise().
     • For PreTrainedTokenizerFast, the .backend_tokenizer.pre_tokenizer and .backend_tokenizer.normalizer fields can both be None, rather than an identity transform like in TkTkT, meaning you always have to check if they exist. Also funny: even when they are not None, you can't check if they exist with a simple if t.backend_tokenizer.normalizer: ... because somehow that's always False.
   • Also, the PreTrainedTokenizerBase interface is not defined with @abstractmethod but with an ever-increasing amount of raise NotImplementedError methods. In other words: it's hard to know which methods need to be implemented and there's no enforcement mechanism to ensure everything has been implemented.
6. Some preprocessors are literally pickled (i.e. the Python object in memory is converted to bytes), encoded in base64, and saved to tokenizer.json. The result is a completely opaque black-box preprocessor (which tokenizers just calls "Precompiled").
   • You have no access to its internals in Python, it has no interface, and without the original training script and the exact versions of all the packages involved, you will never find a description of what it does. This is probably a holdover from adapting the sentencepiece package to HuggingFace, but it doesn't have to be that way. TkTkT has transparent preprocessors.
7. The tokenizers.pre_tokenizers submodule has weird technical debt.
   • The one way to apply GPT-2's mapping from Unicode codepoints to UTF-8 bytes is to use the ByteLevel pretokeniser. What does it do? Much more than just the GPT-2 mapping: it first splits the text in front of each spaces and punctuation mark (optionally prefixed by a space). This is wrong for at least three reasons:
     • Users of the byte mapping don't necessary want the string to be split;
     • It synonymises prefixed spaces (converted to Ġ) with start-of-word boundaries whilst actually all words (even those directly preceded by punctuation) should be marked with such a boundary;
     • It assumes that such boundaries should always be at the start of a word.
   • The GPT-2 convention of having a word boundary at the start of (almost) all words is hardcoded throughout transformers and tokenizers (with options that commonly look like add_prefix_space) even though the original BPE paper used word boundaries at the end of words (</w>). Only supporting the start-of-word convention is bad because this deteriorates downstream performance for e.g. Germanic languages, where a compound has its head at the end and hence it should be allowed to tokenise the head with the exact same tokens as it would be if it was isolated.
8. In the little HuggingFace tokeniser documentation that does exist, there are so many theoretical inaccuracies (e.g. for WordPiece and KudoPiece/ULM) that we shouldn't even have confidence in anything that isn't a BPE tokeniser implemented by them. Their explanation for KudoPiece, an algorithm which itself was already poorly explained originally, is mathematically absurd.

There is also the pyonmttok package which has better design than tokenizers, but also sticks to BPE and KudoPiece.

Pronunciation

The acronym stands for ToKeniser ToolKiT and is supposed to be pronounced fast (kind of like "tuh-kuh-tuh-kuh-ts" but as fast as you can). It is mandatory that you do this.

If you are Brazilian, you may pronounce it "tuca tuca" while playing the official TkTkT theme song (yes, the demented state of modern-day tokeniser implementations will leave you with an equally demented taste in music; but I dance salsa on the weekends).