Skip to main content

Afrisyl-tokenizer

A syllable-aware tokenizer for African languages

Afrisyl-tokenizer is built to fix a core problem: BPE and WordPiece tokenizers break African languages. Instead of splitting "ndinoda" into ['n', 'din', 'oda'], we respect the natural CV syllable structure of Bantu languages.

Built as part of MSc research by Nkosilomusa Ncube, Alumni WeThinkCode.

Key Features

  • Syllable-based: Uses CV, CVC, V patterns found in Shona, Ndebele, Zulu, Swahili, etc
  • Low fertility: 30% fewer tokens than BPE on Shona text
  • Fast: Pure Python + regex. No dependencies for inference
  • HF Compatible: Drop-in replacement for HuggingFace tokenizers
  • Trainable: Train your own tokenizer on any African language corpus
  • Small vocab: 8k-16k vocab covers 95%+ of tokens
  • Open Source: MIT Licensed for research and social impact

Installation

pip install afrisyl-tokenizer

Quick Usage

1. Load Tokenizer

from afrisyl_tokenizer import AfriSylTokenizer

# Load default shona vocab from package
tok = AfriSylTokenizer(language="shona")

# Or load custom vocab
# tok = AfriSylTokenizer(vocab_path="path/to/ndebele_vocab.json")

# Batch processing
texts = ["ndinoda rubatsiro", "mhoroi shamwari"]

batch = tok.batch_encode(
    texts, 
    add_bos=True, 
    add_eos=True, 
    max_len=32, 
    padding=True
)
# batch['input_ids'] -> [[1, 12, 45, ...], [1, 8, 22, ...]]
# batch['attention_mask'] -> [[1, 1, 1, ...], [1, 1, 1, ...]]

decoded_batch = tok.batch_decode(batch["input_ids"])
print(decoded_batch)
# ['ndinoda rubatsiro', 'mhoroi shamwari']

Whitespace handling

Spaces are tokenized the same way o200k (GPT-4o's tokenizer) handles them: a space is fused onto the front of the token that follows it (e.g. " ya") instead of being dropped or turned into its own token between every word. decode(encode(text)) is fully reversible, spacing included:

tok = AfriSylTokenizer(language="shona")
ids = tok.encode("nyika yakatanga")
tok.decode(ids) == "nyika yakatanga"  # True

Every syllable, fallback letter, digit and punctuation mark also has a space-prefixed sibling in the vocab ("ba" and " ba" are both real tokens), and multi-space/newline runs get their own tokens too -- the same trick o200k uses so whitespace doesn't need special-casing at decode time.

Emoji & other scripts

Anything outside the vocab -- emoji, accented letters, other scripts -- falls back to its raw UTF-8 bytes (256 <0xXX> byte tokens) instead of a lossy <unk>, the same way byte-level BPE tokenizers like o200k always have full coverage. This also correctly round-trips multi-codepoint emoji sequences (ZWJ family emoji, flags, skin-tone modifiers):

tok = AfriSylTokenizer(language="shona")
ids = tok.encode("mhoroi shamwari 😀🇿🇼")
tok.decode(ids) == "mhoroi shamwari 😀🇿🇼"  # True

Metadata

Release files for afrisyl-tokenizer 0.1.2

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for afrisyl-tokenizer 0.1.2
File Size Uploaded Provenance Actions
afrisyl_tokenizer-0.1.2.tar.gz 49.0 kB Without provenance Details for afrisyl_tokenizer-0.1.2.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for afrisyl-tokenizer 0.1.2
File Interpreter ABI Platform Actions
afrisyl_tokenizer-0.1.2-py3-none-any.whl Python 3 none any Details for afrisyl_tokenizer-0.1.2-py3-none-any.whl

Total release size: 95.8 kB

Release files / afrisyl_tokenizer-0.1.2.tar.gz

Download URL afrisyl_tokenizer-0.1.2.tar.gz
Size 49.0 kB
Tags Source
SHA-256 checksum
How to use checksums
4e86787b52e624084eddf3077c4ea37207c5cf38eb7a374050d9b6f5bafc950a
BLAKE2b-256 checksum
How to use checksums
2d6d263c249db3195472701e96e1350832c5c792a1d97176cb63f2eeb10b5357
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.0

Release files / afrisyl_tokenizer-0.1.2-py3-none-any.whl

Download URL afrisyl_tokenizer-0.1.2-py3-none-any.whl
Size 46.8 kB
Tags Python 3
SHA-256 checksum
How to use checksums
fb57994d7e9c7fd76c5d6c2fef4a97bd8c874fdae18bf790dd51e7aefde472c2
BLAKE2b-256 checksum
How to use checksums
6c5d57b518f26c731ff4b119278e1811d9233b2fb51760d9cd166bf20968cc78
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.0

Security for afrisyl-tokenizer 0.1.2

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page