Get started with LLM training on Modal

This post introduces Modal Dojo, a library designed to simplify production-grade LLM training. It provides building blocks, tuned defaults, and observability for training on custom datasets and tuning training recipes. The library leverages open-source frameworks like Miles and Slime, with contributions upstreamed by Modal. It demonstrates a concise API for model selection, dataset loading (from Hugging Face, Harbor, or production traces), and recipe configuration. The post also highlights the integrated observability dashboard for tracking loss, reward curves, agent trajectories, and substep timing, and emphasizes how agents can utilize the CLI and skills for training tasks.