Training & Pre-training
Months of thousands of GPUs predicting the next token, trillions of times.
You're looking at a concept explainer: what this term means in plain English, and the records it ties together.
Pre-training runs a model over trillions of tokens, nudging weights after every prediction error, in synchronized runs across tens of thousands of accelerators — the workload the biggest clusters on this site were built for. A frontier run costs hundreds of millions of dollars in compute, fails in exotic ways (silent data corruption, stragglers, loss spikes), and is why checkpointing, interconnects and reliability engineering are front-page infrastructure topics.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Related
Sources
Llama 3 herd paper (training at scale) ↗
Record reviewed as of 2026-01-15 — see Methodology.