Tutorials

Step-by-step guides that take you from installation to a working model. Each tutorial is self-contained and builds on the previous one.

Prerequisites

• Python 3.12+

• JAX with CUDA 12: pip install -U "jax[cuda12]"

• DantinoX: pip install -e ".[all]" (from the repository root)

If you are new to JAX, the JAX quickstart covers the key concepts (JIT compilation, functional transforms, device arrays) in about 15 minutes.

---

Available Tutorials

Tutorial	What you will learn
Training Your First Model	Train a character-level AR Transformer on a text corpus, evaluate it, and generate text
LoRA Fine-Tuning	Adapt a pretrained checkpoint to a new domain using LoRA adapters
Masked Diffusion LM	Train a Masked Diffusion Language Model and use it for generation and infilling
Pushing to HuggingFace Hub	Publish a trained checkpoint, load it on any machine, and share it publicly

---

Choosing the Right Starting Point

I want to train a basic model quickly.
→ Start with Training Your First Model. A small GQA model on a local text file finishes in under 10 minutes on a single GPU.

I have a pretrained checkpoint and want to specialise it.
→ Go to LoRA Fine-Tuning. LoRA trains ~0.2 % of parameters, so it is much faster than full fine-tuning.

I want to explore non-autoregressive generation.
→ See Masked Diffusion LM. Diffusion models enable native infilling and bidirectional context.

I want to share my model.
→ See Pushing to HuggingFace Hub. A single dantinox push command packages and uploads the checkpoint.