penberg / noodle

What is noodle?

Noodle lets you train tiny transformer language models from scratch on custom text data, like TinyStories stories or Dolly instructions, using a simple Rust CLI. Fire up `cargo run train` on your corpus to build a base model, then `finetune` for chat-friendly responses, or jump into `chat` for interactive generation—all accelerated by WGPU, CUDA, or CPU. It's a no-fuss way to experiment with LLMs without massive datasets or Python bloat, perfect for noodle small pack experiments on your machine.

Why is it gaining traction?

In a sea of chili tomato noodle github distractions and noodlemaker wiesbaden gimmicks, Noodle stands out with dead-simple Modal cloud GPU scripts for training without local hardware hassles, plus eval perplexity metrics and repetition-penalty sampling in chat. Devs dig the educational transparency—you see exactly what a small noodle making machine for home spits out—while Rust's speed crushes inference on edge devices, unlike sluggish Python alternatives.

Who should use this?

ML tinkerers training toy models on personal corpora, Rust devs embedding lightweight inference in apps, or educators demoing transformers in noodle days mainz workshops. Ideal for fine-tuning small noodle pasta datasets into instruction-followers, not scaling to GPT-scale.

Verdict

With 25 stars and 1.0% credibility, it's raw and docs-light but surprisingly polished for learning—train a coherent storyteller in 20 epochs on a laptop GPU. Grab it if you want a noodle umfrage on from-scratch LLMs; skip for production unless you're extending it.

(198 words)