Large base models are commonly trained in data from many domains, data combination – Part of each of the used domain – playing an important role in model. The common way of choosing this mixture depends on the trial and error, which is an unemployment that is not a great doubt. We propose a systematic way to find the relevant data mixture of any target domain using measurement rules. Our accuracy of speech predicts loss of model Ni trained D Tokens and the weight vector of a particular domain h. We guarantee the cunning laws by showing the power of predicting three different settings and officials: the largest language model (NLM), and larger models. We also reflect that these measuring laws can release new combustions and their scales: Their parameters can be accurate using a few minimum training, and is used to estimate the performance of large scales and the weight of the invisible domain. Equity laws allow for the relevant domains of any regulated background under the provided budget provided (N, d), providing a toxic manner of expensive exams and mistakes.
