Drop-in Rust ports of Python libraries
We rewrite popular pure-Python libraries in Rust and test every output against the original, down to the last bit of every float. Change one import: your scores, languages and rankings stay exactly the same.
rouge-score → rouge-score-rs
30–80×
faster on one core, up to 353× with score_batch
The implementation behind Hugging Face Evaluate, with the same tokenizer, Porter stemmer and rougeLsum. Unicode tokenizers for other languages.
pip install rouge-score-rslangdetect → langdetect-rs
40–52×
faster on one core, 245–315× with detect_batch
Every probability matches, random walk included, in a third of the memory. install() speeds up libraries that import langdetect themselves.
pip install langdetect-rsrank_bm25 → rank-bm25-rs
635–952×
faster per query on 10k–100k documents, 4–6× faster to index
BM25Okapi, BM25L and BM25Plus with identical scores. install() speeds up LangChain's BM25Retriever without changing its code.
pip install rank-bm25-rsHow we check
A faster library that gives slightly different numbers changes every result built on it. So each port is compared with the original without any tolerance, and a single differing float fails the tests.
Read how langdetect was ported, and why Python 3.12 changes its answers →Exact comparison
Every score, probability and ranking is compared with the Python original on real data and random inputs, with no tolerance.
Same API
The same classes, options and return types, so a migration is one import. Known differences are listed on each page.
Reproducible benchmarks
Speed is measured on public datasets with pinned versions, and the commands to rerun it are published.
Libraries
LSTM, GRU and BiLSTM networks with full backpropagation through time, gradients checked against finite differences and PyTorch.
cargo add rust-lstmLong-term memory for AI agents in Ruby: recall the facts that matter and add them to the prompt, stored in your own Postgres.
gem install engramItem response theory in Ruby: Rasch, 2PL and 3PL models with missing-data strategies and adaptive optimisation.
gem install irt_rubyResearch
Each study ships with its code, prompts, raw outputs and a DOI on Zenodo.
Oct 2026 · O. Kholodniak, O. Prokhorov
How dropout, zoneout and gate initialisation change how far gradients reach back in time in LSTM and GRU networks, and whether that reach, measured early in training, predicts which runs learn long-range dependencies. Trained with rust-lstm.
Oct 2026 · O. Kholodniak
Item response models against simple estimators on public model × item correctness matrices: how well each estimates item difficulty, a new model's capability and its full-benchmark accuracy when only 0.5–80% of the items are run.
Oct 2026 · O. Kholodniak
Choosing per request between a small on-device model, a larger LAN server and a cloud API: latency model, routing policies, a netem testbed and the measurement data.
Updates
rank_bm25 with identical scores, up to 950× faster per query, on PyPI and crates.io.
langdetect to Rust bit for bit, and why Python 3.12 changes its answers.
langdetect with identical results, 40–52× faster, on PyPI and crates.io.
rouge-score on one core and up to 350× with score_batch, same scores.
score_batch.
rouge-score-rs on PyPI and the SyntaxSpirits/rouge metric for Hugging Face Evaluate.
seed, cell-update dropout and GRU zoneout.
SyntaxSpirits is maintained by Oleksandr Kholodniak, a software engineer and PhD researcher at the National Aerospace University “Kharkiv Aviation Institute” in Ukraine. The libraries come out of production Ruby work and research on evaluating language models.
Is there a slow Python library you would like ported, or a result that differs? Open an issue on the library's GitHub page, or write to hello@syntaxspirits.dev.