Saved in:
Bibliographic Details
Main Authors: Liu, Emmy, Bertsch, Amanda, Sutawika, Lintang, Tjuatja, Lindia, Fernandes, Patrick, Marinov, Lara, Chen, Michael, Singhal, Shreya, Lawrence, Carolin, Raghunathan, Aditi, Gashteovski, Kiril, Neubig, Graham
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2503.03862
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908857351012352
author Liu, Emmy
Bertsch, Amanda
Sutawika, Lintang
Tjuatja, Lindia
Fernandes, Patrick
Marinov, Lara
Chen, Michael
Singhal, Shreya
Lawrence, Carolin
Raghunathan, Aditi
Gashteovski, Kiril
Neubig, Graham
author_facet Liu, Emmy
Bertsch, Amanda
Sutawika, Lintang
Tjuatja, Lindia
Fernandes, Patrick
Marinov, Lara
Chen, Michael
Singhal, Shreya
Lawrence, Carolin
Raghunathan, Aditi
Gashteovski, Kiril
Neubig, Graham
contents Improvements in language model capabilities are often attributed to increasing model size or training data, but in some cases smaller models trained on curated data or with different architectural decisions can outperform larger ones trained on more tokens. What accounts for this? To quantify the impact of these design choices, we meta-analyze 92 open-source pretrained models across a wide array of scales, including state-of-the-art open-weights models as well as less performant models and those with less conventional design decisions. We find that by incorporating features besides model size and number of training tokens, we can achieve a relative 3-28% increase in ability to predict downstream performance compared with using scale alone. Analysis of model design decisions reveal insights into data composition, such as the trade-off between language and code tasks at 15-25\% code, as well as the better performance of some architectural decisions such as choosing rotary over learned embeddings. Broadly, our framework lays a foundation for more systematic investigation of how model development choices shape final capabilities.
format Preprint
id arxiv_https___arxiv_org_abs_2503_03862
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Not-Just-Scaling Laws: Towards a Better Understanding of the Downstream Impact of Language Model Design Decisions
Liu, Emmy
Bertsch, Amanda
Sutawika, Lintang
Tjuatja, Lindia
Fernandes, Patrick
Marinov, Lara
Chen, Michael
Singhal, Shreya
Lawrence, Carolin
Raghunathan, Aditi
Gashteovski, Kiril
Neubig, Graham
Computation and Language
Artificial Intelligence
Improvements in language model capabilities are often attributed to increasing model size or training data, but in some cases smaller models trained on curated data or with different architectural decisions can outperform larger ones trained on more tokens. What accounts for this? To quantify the impact of these design choices, we meta-analyze 92 open-source pretrained models across a wide array of scales, including state-of-the-art open-weights models as well as less performant models and those with less conventional design decisions. We find that by incorporating features besides model size and number of training tokens, we can achieve a relative 3-28% increase in ability to predict downstream performance compared with using scale alone. Analysis of model design decisions reveal insights into data composition, such as the trade-off between language and code tasks at 15-25\% code, as well as the better performance of some architectural decisions such as choosing rotary over learned embeddings. Broadly, our framework lays a foundation for more systematic investigation of how model development choices shape final capabilities.
title Not-Just-Scaling Laws: Towards a Better Understanding of the Downstream Impact of Language Model Design Decisions
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2503.03862