A Python package for consuming Wikipedia language models including tokenizers, n-gram models, Markov chains, vocabularies, and datasets.

• BPE Tokenizers: Pre-trained tokenizers for 300+ languages and utilities for LLM integration
• N-gram Models: Simple language models for text scoring and next token prediction
• Markov Chains: Text generation models with configurable depth
• Vocabularies: Comprehensive word dictionaries with frequency information
• Embeddings: Position-aware cross-lingual word embeddings via BabelVec designed for resource-constrained environments
• Datasets: Wikipedia text data in various splits (1k, 5k, 10k, train) available via wikisets.
• Multi-language Support: All models available for 300+ Wikipedia languages
• Comprehensive Evaluation: Each language includes complete model evaluations and metrics available at https://huggingface.co/wikilangs/{lang} (e.g., https://huggingface.co/wikilangs/ary, https://huggingface.co/wikilangs/en, https://huggingface.co/wikilangs/fr)
• Easy API: Simple, intuitive interface for loading and using models

pip install wikilangs

from wikilangs import tokenizer, ngram, markov, vocabulary, embeddings, languages

# Create a tokenizer (date defaults to 'latest')
tok = tokenizer(lang='en', vocab_size=16000)

# Tokenize text
tokens = tok.tokenize("Hello, world!")
token_ids = tok.encode("Hello, world!")
print(tokens)  # ['_he', 'l', 'lo', ',', '_world', '!']
print(token_ids)  # [1234, 5, 5678, 9, 10, 11]

# Create an n-gram model
ng = ngram(lang='en', gram_size=3)

# Score text
score = ng.score("This is a sample sentence.")
print(score)  # -12.345

# Predict next token
predictions = ng.predict_next("This is a", top_k=5)
print(predictions)  # [('sample', 0.85), ('test', 0.05), ...]

# Create a Markov chain
mc = markov(lang='en', depth=2)

# Generate text
text = mc.generate(length=50)
print(text)  # "Generated text using the Markov chain model..."

# Create a vocabulary
vocab = vocabulary(lang='en')

# Look up a word
word_info = vocab.lookup("example")
print(word_info)  # {'token': 'example', 'frequency': 12345, 'idf_score': 7.91, 'rank': 25436}

# Create embeddings
emb = embeddings(lang='ary', dimension=32)

# Get word vector
vec = emb.embed_word("مرحبا")
print(vec.shape)  # (32,)

# Get sentence vector (supports average, rope, decay, sinusoidal)
sent_vec = emb.embed_sentence("مرحبا بالعالم", method='rope')
print(sent_vec.shape)  # (32,)

# List available languages
available_langs = languages()
print(f"Available languages: {available_langs[:5]}...")

Create a BPE tokenizer instance.

Parameters:

• lang (str): Language code (e.g., 'en', 'fr', 'ary')
• date (str): Date of the model (format: YYYYMMDD, default: 'latest')
• vocab_size (int): Vocabulary size (8000, 16000, 32000, 64000)
• format (str): Output format ('sentencepiece' or 'huggingface')

Returns:

• BPETokenizer: Initialized tokenizer instance

Create an n-gram model instance.

Parameters:

• lang (str): Language code (e.g., 'en', 'fr', 'ary')
• date (str): Date of the model (format: YYYYMMDD, default: 'latest')
• gram_size (int): Size of n-grams (2, 3, 4, 5)
• variant (str): Type of n-grams ('word' or 'subword')

Returns:

• NGramModel: Initialized n-gram model instance

Create a Markov chain model instance.

Parameters:

• lang (str): Language code (e.g., 'en', 'fr', 'ary')
• date (str): Date of the model (format: YYYYMMDD, default: 'latest')
• depth (int): Depth of the Markov chain (1, 2, 3, 4, 5)
• variant (str): Type of transitions ('word' or 'subword')

Returns:

• MarkovChain: Initialized Markov chain model instance

Create a vocabulary instance.

Parameters:

• lang (str): Language code (e.g., 'en', 'fr', 'ary')
• date (str): Date of the model (format: YYYYMMDD, default: 'latest')

Returns:

• WikilangsVocabulary: Initialized vocabulary instance

Create an embeddings instance using BabelVec.

Parameters:

• lang (str): Language code (e.g., 'en', 'fr', 'ary')
• date (str): Date of the model (format: YYYYMMDD, default: 'latest')
• dimension (int): Embedding dimension (default: 32)

Returns:

• Embeddings: Initialized embeddings instance (if babelvec is installed)
• tuple: (file_path, metadata) if babelvec is not installed

List available language codes for a given date.

Parameters:

• date (str): Date of the dataset (format: YYYYMMDD, default: 'latest')

Returns:

• list[str]: List of available language codes

Get available language codes with ISO 639 metadata enrichment.

Parameters:

• date (str): Date of the dataset (format: YYYYMMDD, default: 'latest')

Returns:

• list[LanguageInfo]: List of LanguageInfo objects with ISO 639 metadata (name, alpha_2, alpha_3, etc.)

Models are available for 300+ Wikipedia languages including:

• English (en)
• French (fr)
• Spanish (es)
• German (de)
• Arabic (ar)
• Chinese (zh)
• Japanese (ja)
• Korean (ko)
• And many more...

Models are updated regularly. Check the Hugging Face organization for the latest available dates.

For advanced embedding operations, install BabelVec:

pip install babelvec

Using the wikilangs API:

from wikilangs import embeddings

# Load embeddings (defaults to 32 dimensions)
emb = embeddings(lang='ary')

# Get word vector
vec = emb.embed_word("مرحبا")

# Position-aware sentence embedding (supports 'average', 'rope', 'decay', 'sinusoidal')
sent_vec = emb.embed_sentence("مرحبا بالعالم", method='rope')

Or using BabelVec directly:

from huggingface_hub import hf_hub_download
from babelvec import BabelVec

# Load embeddings
embedding_file = hf_hub_download(
    repo_id='wikilangs/ary',
    filename='models/embeddings/monolingual/ary_32d.bin',
    repo_type='model'
)
model = BabelVec.load(embedding_file)

Check out the demo scripts:

• demo_models.py - Basic model operations
• demo_embeddings.py - Embedding operations with BabelVec
• demo_comprehensive.py - All models working together

pip install -r requirements.txt

pytest tests/

We are deeply grateful to our generous sponsor Featherless.ai for making this project possible.

Created and maintained by Omar Kamali from Omneity Labs.

Wikilangs is built on top of the incredible work by the Wikimedia Foundation and the open-source community. All content maintains the original CC-BY-SA-4.0 license.

MIT License - see LICENSE for details.

If you use this package in your research, please cite:

@misc{wikilangs2025,
  title = {Wikilangs: Open NLP Models for Wikipedia Languages},
  author = {Kamali, Omar},
  year = {2025},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18073153},
  url = {https://huggingface.co/wikilangs},
  institution = {Omneity Labs}
}