🏷️ gliclass-rs: GLiClass inferences in Rust

💬 Introduction

gliclass-rs is an inference engine for GLiClass language models.

These models are efficient for zero-shot topic classification or derivatives like sentiment analysis. They can also be used for efficient re-ranking.

"GLiClass" stands for "Generalist and Lightweight Model for Sequence Classification", after an original work from Knowledgator, which was itself inspired by GLiNER.

gliclass-rs is built in pure Rust, as an 🧩 orp pipeline.

🎓 Examples

[dependencies]
"gliclass-rs" = "0.9.x"

use gliclass::{GLiClass, params::Parameters, input::text::TextInput};

let gliclass = GLiClass::new("tokenizer.json", "model.onnx", Parameters::default())?;
        
let input = TextInput::from_str(
    &[
        "Rust is a systems programming language focused on safety, speed, and concurrency.",
        "Traveling is the perfect way to explore new cultures through their food.",
        "Traveling for science allows researchers to collaborate with experts worldwide.",
    ],
    &["computing", "science", "programming", "travel", "food", "politics"]
);    

let classes = gliclass.inference(input)?;

for i in 0..classes.len() {
    println!("Text {i}: {}", classes.best_label(i, None).unwrap());        
}

Please refer the the source code in examples for complete examples.

🧬 Models

Currently gliclass-rs has been tested with the following models:

Model	Download	prompt_first
gliclass-small-1.0	HF Hub	false
gliclass-large-1.0	HF Hub	false
gliclass-modern-base-v2.0-init	HF Hub	true
gliclass-modern-large-v2.0	HF Hub	true

It should work out-of-the-box with other equivalent models, please report your own experiments.

⚠️ Take care of setting the prompt_first parameter according the selected model's expectations, or read parameters from the config.json that goes with the model (0.9.1+).

👉 Related

This project follows the same principles as the ones below. Refer to their documentation for more details:

• 🌿 gline-rs: inference engine for GLiNER models
• 🧲 gte-rs: general text embedding and re-ranking