Chinchilla scaling
DeepMind's 2022 finding that, for a fixed budget, models should be trained on about 20 tokens per parameter.
Also: compute-optimal training, Chinchilla law
Jordan Hoffmann and colleagues at DeepMind trained over 400 models in 2022 and concluded that earlier large models were too big for the data they saw. For a given compute budget, parameters and training tokens should grow in equal proportion. Their 70-billion-parameter Chinchilla, trained on 1.4 trillion tokens, beat the 280-billion-parameter Gopher. Later practice went further: because smaller models are cheaper to run, labs now train them far past the Chinchilla ratio, as with Llama 3 8B on over 15 trillion tokens.