The Smol Training Playbook

A practical, opinionated guide to fine-tuning small language models, built by the HuggingFace team that actually ships SmolLM. This is not a paper or a blog post — it is a runnable playbook that walks through data prep, training configs, and evaluation for models you can actually deploy on cheap hardware. What makes it worth your Saturday: it is written for people who want a working model at the end, not a PhD thesis. The SmolLM team has strong opinions about what cuts matter and what is cargo-culted from larger-model recipes, and those opinions are baked into the defaults here. Reservation: if you are already fine-tuning on a regular basis, some of this will be familiar. The real value is in the curated shortcuts for people doing this for the first time or the second time. -> Best for: AI engineer or solo founder who wants to ship a fine-tuned model without a GPU cluster