Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Scaling Laws for Neural Language Models

About

We study empirical scaling laws for language model performance on the cross-entropy loss. The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude. Other architectural details such as network width or depth have minimal effects within a wide range. Simple equations govern the dependence of overfitting on model/dataset size and the dependence of training speed on model size. These relationships allow us to determine the optimal allocation of a fixed compute budget. Larger models are significantly more sample-efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data and stopping significantly before convergence.

Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, Dario Amodei• 2020

Related benchmarks

TaskDatasetResultRank
Image GenerationImageNet-1k (val)
FID151.9
106
Multimodal Question AnsweringScienceQA--
61
Language ModelingC4 (train)
PPL19.21
50
Language ModelingLAMBADA zero-shot (test)
Accuracy (zero-shot)57.23
44
Language ModelingWikiText-103 zero-shot (test)
PPL14.14
34
Zero-shot EvaluationAI2 OLMES zero-shot--
15
Scaling Law ModelingPythia AWQ 4-bit
R2 Score0.9764
8
Scaling Law ModelingPythia bnb 4-bit
R2 Score96.79
8
Model ExtrapolationPythia k=3 (1B, 410M, 160M)
Pooled R^2-0.507
8
Model ExtrapolationPythia k=4 (≤2.8B)
Pooled R^2-0.048
8
Showing 10 of 22 rows

Other info

Follow for update