Researchers have introduced new scaling laws called Compute-Data (CD) scaling laws, which offer a unified framework for pretraining large language models. These laws bridge the gap between compute-optimal scaling, where data is abundant, and data-optimal scaling, where the dataset is fixed. The framework incorporates a token-effectiveness function that quantifies the value of derived tokens relative to fresh ones, considering factors like multi-epoch repetition and paraphrasing. This function reveals that token effectiveness is not constant but depends on model size, the ratio of tokens to parameters, and the amount of derived data, suggesting that classical compute-optimal allocation is suboptimal in most practical scenarios. AI
IMPACT Introduces a more nuanced approach to LLM pretraining resource allocation, potentially optimizing future model development.
RANK_REASON The cluster contains a research paper detailing new theoretical scaling laws for LLM pretraining. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
