FastText.Net

A managed, dependency-free C# port of Facebook fastText inference, focused on text/language classification. Load a pretrained supervised model (such as lid.176.ftz for language identification) and classify text without shipping any native binaries.

This is a port of the prediction path only (model loading + predict); training is not included. See THIRD-PARTY-NOTICES.md for attribution.

Unofficial. This is an independent, community port of the now-archived fastText project. It is not affiliated with or endorsed by Meta / Facebook.

Features

• Pure C# (net10.0), no native dependency, no P/Invoke.
• Reads the standard fastText binary model format, including quantized .ftz models (product quantization) and hierarchical-softmax / softmax / sigmoid losses.
• Byte-exact tokenization, FNV-1a hashing, and subword n-grams — results match the reference implementation (verified against the official Python fasttext package).
• SIMD-accelerated linear algebra via System.Numerics.Tensors (TensorPrimitives).
• Allocation-light prediction hot path (pooled buffers, thread-local state).

Usage

using FastTextNet;

var model = FastTextModel.Load("lid.176.ftz");

foreach (var p in model.Predict("Bonjour, comment allez-vous?", k: 3))
{
    Console.WriteLine($"{p.Label}: {p.Probability:P2}");
}
// __label__fr: 95.79%
// __label__en:  1.89%
// __label__nl:  0.26%

FastTextModel is immutable after loading and safe to share across threads.

Getting a model

Language-identification models are published by Facebook (CC BY-SA 3.0) and are not bundled here:

• lid.176.ftz (~917 KB, quantized): https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.ftz
• lid.176.bin (~126 MB, full): https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.bin

Layout

Path	Description
src/FastText.Net/	The library.
tests/FastText.Net.Tests/	xUnit correctness tests vs. the reference fasttext outputs.
bench/FastText.Net.Benchmarks/	BenchmarkDotNet performance suite.

Building & testing

dotnet build -c Release
dotnet test  -c Release           # place lid.176.ftz under tests/.../models or set FASTTEXT_LID_MODEL
dotnet run   -c Release --project bench/FastText.Net.Benchmarks

Benchmarks

Measured on the quantized lid.176.ftz model (16-dim embeddings, 176 labels) on a single thread. See bench/results.md for the full report and the machine it was captured on.

License

MIT — see LICENSE. Derived from fastText (MIT, © Facebook, Inc.).