This might seem like an odd way to start an “LLM training guide”. But many failed training projects didn’t fail because of bad hyperparameters or buggy code, they failed because someone decided to train a model they didn’t need. So before you commit to training, and dive into how to execute it, you need to answer two questions: why are you training this model? And what model should you train? Without clear answers, you’ll waste months of compute and engineering time building something the world already has, or worse, something nobody needs
Here’s a simple test: spend a few days building on top of Qwen3, Gemma3, or another current SOTA model. Can you reach your performance goals through prompting, tool-use, or post-training? If not, it’s probably time to train your own.
A change is derisked when testing shows it either improves performance on your target capabilities, or provides a meaningful benefit (e.g. faster inference, lower memory, better stability) without hurting performance beyond your acceptable tradeoffs.
