Skeletoken

This package contains Pydantic datamodels that fully describe the tokenizer.json file used in transformers via Tokenizers. This is useful, because working with this format is complicated.

Rationale

In one sentence: Validate, edit, and transform Hugging Face tokenizers safely.

The Hugging Face tokenizers representation does not reliably allow you to edit tokenizers as a structured object. This means that complex changes to tokenizers require you to edit the tokenizer.json file manually. This is annoying, because the format of this file is complicated.

Furthermore, tokenizers does not give reasonable errors when parsing a tokenizer fails. It does give line/character numbers, but those point to the last character of the section where the parsing fails. For example, inserting an illegal vocabulary item just tells you that there is an issue in the vocabulary somewhere by pointing out the last character of the vocabulary as the place where the error occurs.

This package contains datamodels (pydantic datamodels) that contain the same constraints as the tokenizers package. In other words, if you can create a model in this package, the tokenizers package can parse it. This allows you to progressively edit tokenizer json files, all the while getting productive error messages.

Installation

Install it via pip

pip install skeletoken

What can it do?

skeletoken allows you to:

• validate tokenizer.json with human-readable errors
• edit tokenizers as typed objects (Pydantic), catching invalid values at construction time instead of a cryptic tokenizers parse error later
• safely add and remove tokens from vocabularies without breaking anything
• consolidate vocabularies, merging duplicate/redundant entries while keeping the merge table and IDs consistent
• apply common transformations (decasing, greedy merges, etc.)
• auto-fix common inconsistencies, e.g. adding special tokens referenced by a post-processor but missing from the vocabulary, or pruning stale AddedToken entries
• round-trip to tokenizers and transformers without silently dropping fields
• apply tokenization changes to transformers, sentence-transformers, pylate and model2vec models, resizing and remapping the embedding table so tokens that survive an edit keep their learned embeddings

Example

Here's some examples of what skeletoken can do:

Autofixing a tokenizer

skeletoken autofixes any tokenizer you load. See automatic checks to see what gets fixed automatically. For example, the Qwen/Qwen3-0.6B tokenizer has a lot of special tokens that are not part of the regular tokenizer vocabulary. This leads to a mismatch between the size of a tokenizer and the number of tokens that tokenizer can produce. skeletoken adds these to the vocabulary automatically.

from transformers import AutoTokenizer
from skeletoken import TokenizerModel

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
# Mismatch due to missing special tokens
print(tokenizer.vocab_size)  # 151643
print(len(tokenizer))  # 151669

# Load a model from the hub.
tokenizer_model = TokenizerModel.from_pretrained("Qwen/Qwen3-0.6B")
# Convert the tokenizer to transformers
tokenizer = tokenizer_model.to_transformers()
# All missing special tokens have been added to the vocabulary
print(tokenizer.vocab_size)  # 151669
print(len(tokenizer))  # 151669

Adding components to a tokenizer

skeletoken can add components to a tokenizer. First we load one, and inspect it:

from skeletoken import TokenizerModel

# Directly pull a tokenizer from the hub
tokenizer_model = TokenizerModel.from_pretrained("gpt2")

print(tokenizer_model.model.type)
# ModelType.BPE
print(tokenizer_model.pre_tokenizer.type)
# PreTokenizerType.BYTELEVEL

We can then add a digit splitter to the tokenizer.

from skeletoken import TokenizerModel
from skeletoken.pre_tokenizers import DigitsPreTokenizer

model = TokenizerModel.from_pretrained("gpt2")
tok = model.to_tokenizer()

# Create the digits pretokenizer
digits = DigitsPreTokenizer(individual_digits=True)
model = model.add_pre_tokenizer(digits)

new_tok = model.to_tokenizer()
print(tok.encode("hello 123").tokens)
# ['hello', 'Ġ123']
print(new_tok.encode("hello 123").tokens)
# ['hello', 'Ġ', '1', '2', '3']

Decasing a tokenizer

For background, see this blogpost. Decasing is super easy using skeletoken.

from tokenizers import Tokenizer
from skeletoken import TokenizerModel

model_name = "intfloat/multilingual-e5-small"

tokenizer = Tokenizer.from_pretrained(model_name)

print([tokenizer.encode(x).tokens for x in ["Amsterdam", "amsterdam"]])
# [['<s>', '▁Amsterdam', '</s>'], ['<s>', '▁am', 'ster', 'dam', '</s>']]

model = TokenizerModel.from_pretrained(model_name)
model = model.decase_vocabulary()

lower_tokenizer = model.to_tokenizer()
print([lower_tokenizer.encode(x).tokens for x in ["Amsterdam", "amsterdam"]])
# [['<s>', '▁amsterdam', '</s>'], ['<s>', '▁amsterdam', '</s>']]

Making a tokenizer greedy

For background, see this blog post. Like decasing, turning any tokenizer into a greedy one is super easy using skeletoken.

from tokenizers import Tokenizer
from skeletoken import TokenizerModel

model_name = "gpt2"

tokenizer = Tokenizer.from_pretrained(model_name)

print([tokenizer.encode(x).tokens for x in [" hellooo", " bluetooth"]])
# [['Ġhell', 'ooo'], ['Ġblu', 'etooth']]

model = TokenizerModel.from_pretrained(model_name)
model = model.make_model_greedy()
greedy_tokenizer = model.to_tokenizer()
print([greedy_tokenizer.encode(x).tokens for x in [" hellooo", " bluetooth"]])
# [['Ġhello', 'oo'], ['Ġblue', 'too', 'th']]

Consolidating a vocabulary

Adding a normalizer or pre-tokenizer can leave a vocabulary with entries that have become unreachable, or that now collide with another entry. consolidate_vocabulary finds and removes these automatically, keeping the merge table and IDs consistent.

from skeletoken import TokenizerModel
from skeletoken.normalizers import ReplaceNormalizer

model = TokenizerModel.from_pretrained("bert-base-cased")
# "A" and "a" are currently distinct tokens.
print("A" in model.vocabulary, "a" in model.vocabulary)
# True True

# Add a normalizer that maps "A" to "a", making the two tokens collide.
model = model.add_normalizer(ReplaceNormalizer(pattern="A", content="a"))
model = model.consolidate_vocabulary()

# The now-unreachable duplicate has been removed.
print("A" in model.vocabulary, "a" in model.vocabulary)
# False True

Safely resizing model embeddings

Editing a vocabulary changes token IDs, which normally means your model's embedding table and your tokenizer silently drift out of sync. skeletoken computes a ModelDelta (via TokenizerModel.model_delta) that maps old token IDs to new ones, and uses it to reshape a model's embedding table so that tokens which survive an edit keep their learned embeddings; only genuinely new tokens get fresh rows. This is supported for transformers, sentence-transformers, pylate and model2vec models, so you can prune or edit a vocabulary without retraining from scratch.

from transformers import AutoModel
from skeletoken import TokenizerModel
from skeletoken.external.transformers import reshape_embeddings

model_name = "bert-base-cased"
model = AutoModel.from_pretrained(model_name)
tokenizer_model = TokenizerModel.from_pretrained(model_name)

# Decasing removes duplicate (now-colliding) tokens from the vocabulary.
decased = tokenizer_model.decase_vocabulary()

# The embedding table is resized to match, and every surviving token
# keeps the embedding it already learned.
new_model = reshape_embeddings(model, decased)
print(model.get_input_embeddings().weight.shape[0])       # original vocabulary_size
print(new_model.get_input_embeddings().weight.shape[0])   # decased.vocabulary_size

Roadmap

Here's a rough roadmap:

• ✅ Add automated lowercasing (see blog)
• ✅ Add vocabulary changes + checks (e.g., check the merge table if a token is added)
• ✅ Add helper functions for adding modules
• ✅ Add secondary constraints (e.g., if an AddedToken refers to a vocabulary item does not exist, we should crash.)
• ✅ Add a front end for the Hugging Face trainer
• ✅ Add automatic model editing
• Consistent tokenizer hashing: instantly know if two tokenizers implement the same thing.
• Add a front end for sentencepiece training.

License

MIT

Author

Stéphan Tulkens

Citation

If you use skeletoken in your work, please cite:

@software{stephan_tulkens_2026_19401859,
  author       = {Stephan Tulkens},
  title        = {skeletoken},
  month        = feb,
  year         = 2026,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.19401859},
  url          = {https://zenodo.org/records/19401859},
}