论文卡-ChinchillaPaper Card - Chinchilla

研究问题：摘要原话——给定算力预算下，最优的模型大小与训练 token 数各是多少；结论是「当前大模型普遍训练不足」（"significantly undertrained"），因为业界只顾加大模型、不增加数据。

核心结论原话："for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled"——模型与数据要等比例扩。

证据规模：训练 400 多个模型（70M 到 16B+ 参数、5 到 500B token）做定标拟合，再用 70B 的 Chinchilla 验证——它「超过 Gopher 及其他大模型」（第 4.2 节，当日核对）。

作者自陈的边界（当日核对）：Chinchilla 比 Gopher 多训 4 倍数据，作者自己提醒「训练/测试集泄漏可能人为抬高结果」，因此更看重其他任务——连定标论文都在防数据污染。

它没有主张的：结论来自 ≤16B 模型的拟合再外推——「70B 之后规律不变」是外推不是实测；「参数量不重要」也不是论文说法，是等比例扩。

Research question (abstract): under a fixed compute budget, what is the optimal model size and token count; the finding is that current LLMs are "significantly undertrained" because the field scaled models while keeping data constant.

Conclusion verbatim: "for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled".

Evidence scale: 400+ models trained (70M to 16B+ parameters, 5B to 500B tokens) to fit the scaling law, then validated with the 70B Chinchilla, which "outperforms Gopher" (Section 4.2, verified).

Author-flagged boundary (verified): Chinchilla trained on 4× more data than Gopher, and the authors warn train/test leakage may artificially boost results — even a calibration paper guards against contamination.

What it does not claim: the law is fitted on ≤16B models and extrapolated — 'the rule holds past 70B' is extrapolation, not measurement; 'parameter count doesn't matter' is not the paper's claim either.