v0.1.0 · MIT
word2vec-rs
Production-grade Word2Vec in Rust — Skip-gram & CBOW with negative sampling, subsampling, linear LR decay, plots, and a CLI. Zero Python.
2
architectures
8
modules
45
tests
2
CLIs
API documentation
Full rustdoc — every module, struct, and function with inline examples and doc tests.
Training loss curve
Per-epoch average loss with linear LR decay from 0.025 → 0.0001.
Word vector PCA
2D PCA projection of the top-50 vocabulary words after training.
Training history
JSON — loss, learning rate, pairs processed, and elapsed time per epoch.
Latest training run
Loss vs epoch — Skip-gram, dim=50, 15 epochs
PCA projection — top 50 words
Modules
error
Word2VecError — unified error type via thiserror
config
Config, ModelType — all hyperparameters with defaults
vocab
Frequency counts, min_count, subsampling, unigram noise table
model
Xavier init, skip-gram & CBOW SGD in-place update
trainer
Training loop, linear LR decay, progress bar, epoch history
embeddings
most_similar, similarity, analogy, save/load JSON + text
plot
Loss curve PNG + PCA scatter plot via plotters
bin/train
CLI — --input --output --dim --epochs --model --window