Bashkir FastText Word Embeddings (300d, Mini Beta)

Compact 17 MB subword word embeddings for Bashkir, built for semantic search, lexical tools and on-device NLP.

Overview

Compact subword word embeddings for Bashkir (ba), trained on a monolingual Bashkir-language dataset. The release ships a single quantized model (bashkir_fasttext_300d_mini.bin) intended for semantic search, lexical tools and lightweight NLP experiments on CPU, mobile and edge devices.

At a glance
Task	Feature extraction / word embeddings
Default artifact	`bashkir_fasttext_300d_mini.bin`
Source	A monolingual Bashkir-language dataset
Version / license	v0.1-beta / Apache-2.0

Files and Configurations

File	Description
`bashkir_fasttext_300d_mini.bin`	Quantized 300-dimensional FastText model; recommended for most users
`config.json`	Architecture and runtime parameters
`META.json`	Release passport: version, size, source and artifact hashes
`SHA256SUMS`	SHA-256 checksums for public release files

The full model and text-vector export are not included in this release.

Model Architecture

Property	Value
Architecture	FastText Skip-gram with character subwords
Embedding dimension	300
Character n-grams	3–6
Active vocabulary	top 50,000 words and top 100,000 subword n-grams
Compression	Product Quantization with FP16 codebooks
Model size	17.01 MB
OOV support	vectors synthesized for unseen word forms and suffixes
Format	`compress-fasttext` / Gensim-compatible
Runtime	CPU; no GPU required

Examples

Subword training lets the model connect a root with its inflected forms, for example мәктәп → мәктәптәр → мәктәптәребеҙҙә, and return semantic neighbours for words such as матур, атай, тел and ҡояш, even when a form was never seen during training.

Method

monolingual Bashkir text → Unicode NFC tokenization → skip-gram 300d (char n-grams 3–6)
    → prune (top-50k words, top-100k subwords) → PQ FP16 → mini .bin

The model was trained from a monolingual Bashkir-language dataset. Corpus size, source reference and training parameters are recorded in config.json. Text preparation uses normalized Unicode tokens and preserves Bashkir-specific letters. This release is a compact derivative of the full 300-dimensional model: it retains subword information while reducing the vocabulary and model footprint for local, mobile and edge use.

Quality and Use

This is a beta diagnostic release, not a standardized linguistic benchmark. Independent stress tests showed strong morphology and OOV behaviour, and the full and compact variants produced near-identical results. The official Facebook Bashkir vectors are a useful general-purpose baseline: facebook/fasttext-ba-vectors.

Limitations

Large web-derived corpora may contain borrowed vocabulary, foreign names, technical terms, OCR errors and concatenated tokens.
Raw nearest-neighbour lookups can occasionally return concatenated headlines or compound hyphenations from the source text.
Similarity scores are embeddings-based, not human judgments.
Nearest-neighbour results should be validated before use in a dictionary, spellchecker or production pipeline; combine the model with frequency data, canonical word lists and language filtering where appropriate.

Related Resources

Bashkir Multiclass LID — gate mixed Bashkir/Tatar/Russian input before embedding-based processing.
Bashkir Frequency Index — canonical word frequencies to validate nearest neighbours and rank spellcheck candidates.

Usage

pip install compress-fasttext huggingface_hub

from huggingface_hub import hf_hub_download
from compress_fasttext.models import CompressedFastTextKeyedVectors

path = hf_hub_download(
    repo_id="failed09/bashkir-fasttext",
    filename="bashkir_fasttext_300d_mini.bin",
    revision="main",  # pin a reviewed commit for reproducible deployments
)
model = CompressedFastTextKeyedVectors.load(path)

vector = model.get_vector("мәктәптәребеҙҙә")
print(vector.shape)  # (300,)
print(model.most_similar("матур", topn=5))
print(model.similarity("атай", "әсәй"))

License

The model and release code are provided under the Apache-2.0 license. The model is derived from upstream and project corpora; this release does not claim ownership of source texts or other training materials. Upstream rights and licensing conditions remain applicable.

Citation

@software{failed09_bashkir_fasttext_embeddings_2026,
  title = {Bashkir FastText Word Embeddings},
  author = {failed09},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/failed09/bashkir-fasttext},
  note = {Compact beta subword embeddings for Bashkir}
}

Open Bashkir Data and Sources 🐝

This release is part of an open-source effort to support the development, preservation and practical use of the Bashkir language. Other related models, datasets and tools are available on the author's Hugging Face profile.

The author does not claim ownership or authorship of the source texts or other materials used to derive this release; rights and licensing remain with the original authors, publishers and dataset providers. Source texts are not redistributed in this repository, so users should follow the licenses and attribution requirements of the relevant upstream resources.

Downloads last month: -

failed09
/

bashkir-fasttext