Training Compute-Optimal LLMs — Chinchilla (2022)
研究论文
规模化
DeepMind's correction: most models were undertrained on data.
您正在查看一条研究论文记录：一篇里程碑式的论文，以及它为该领域带来了什么。
Hoffmann et al. showed compute-optimal training wants ~20 tokens per parameter — far more data than the field was using — so a 70B model trained right beat a 280B one. It reshaped every subsequent training run and made data supply a first-order constraint.
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。
基础来源
相关
来源
记录审阅截至 2026-01-15 — 查看 方法论.