CoChat

Scaling Data-Constrained Language Models

Open access

The DOI matches its Crossref record; 149 of the 149 references it cites were found in a registry or an index; 37 of the 48 sources its claims cite were checked for support.

Abstract

The current trend of scaling language models involves increasing both parameter count and training dataset size. Extrapolating this trend suggests that training dataset size may soon be limited by the amount of text data available on the internet. Motivated by this limit, we investigate scaling language models in data-constrained regimes. Specifically, we run a large set of experiments varying the extent of data repetition and compute budget, ranging up to 900 billion training tokens and 9 billion parameter models. We find that with constrained data for a fixed compute budget, training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data. However, with more repetition, the value of adding compute eventually decays to zero. We propose and empirically validate a scaling law for compute optimality that accounts for the decreasing value of repeated tokens and excess parameters. Finally, we experiment with approaches mitigating data scarcity, including augmenting the training dataset with code data or removing commonly used filters. Models and datasets from our 400 training runs are freely available at https://github.com/huggingface/datablations.

Overview

CoChat Summary

TL;DR

The paper studies scaling language models when unique training data is limited, training more than 400 models up to 9B parameters and 900B tokens. It finds that up to 4 epochs of repeated data is nearly as good as unique data, but returns diminish sharply around 16 epochs, and proposes a data-constrained scaling law.

Key findings

  • With constrained data and fixed compute, training up to 4 epochs yields negligible loss change versus unique data; an 8.7B model trained for 4 epochs on 44B unique tokens finished with only 0.5% higher validation loss than the single-epoch model trained on 178B unique tokens. More repetition eventually decays to zero, with \(R^*_D \approx 15\) (about 16 epochs) as the learned “half-life.”
  • Data-constrained allocation should scale epochs faster than parameters: the fit gives \(R^*_N = 5.309743\) and \(R^*_D = 15.387756\), so excess parameters decay faster than repeated data. A 6.3B data-constrained compute-optimal model with 25B unique tokens achieved better loss and downstream performance than the 8.7B Chinchilla-style model despite having 27% fewer parameters.
  • Complementary strategies can extend data: mixing in Python code provides a 2× increase in effective tokens, and filling up to 50% of data with code (42B tokens) showed no deterioration on natural language tasks. Perplexity filtering was effective, while deduplication did not help on C4 but was effective on noisier OSCAR.
  • Double descent was observed: on 100M tokens, loss increased at 200 epochs before decreasing again; most such examples were removed from the scaling-law fitting.

Methods

  • Trained more than 400 GPT-2 architecture/tokenizer transformer language models from 10M to 9B parameters, for up to 900B total tokens and up to 1500 epochs, on subsets of C4, OSCAR, and Python data from The Stack, with data constraints designed for maximal overlap.
  • Used three protocols: fixed unique data (varying epochs and parameters), fixed FLOPs (varying data budget), and parametric fit. The parametric fit generalizes the Chinchilla law by replacing data and parameter terms with effective data and effective parameters via exponential decay constants \(R^*_D\) and \(R^*_N\), fitted on 182 samples.
  • Evaluated held-out test loss (210M C4 validation tokens) and 19 downstream natural language tasks with 0–5 few-shot exemplars, normalizing scores; training used up to 256 AMD Instinct MI250X GPUs on LUMI.

Limitations

  • The paper states it focuses on repeating the entire unique dataset, not fractions; returns from additional epochs may depend heavily on hyperparameters such as learning rate, dropout, or optimizer; findings are dataset-dependent (C4 and OSCAR) and limited to text data and GPT-style architectures; other data strategies and modalities are left to future work.
  • Evident from the text: the fitting removed double-descent and outlier samples where excess epochs or parameters hurt, and the alternative alpha-beta decay formulation lacked rigorous mathematical justification and mispredicted returns. The downstream evaluation is described as not comprehensive, and the scaling laws fitted on C4 were applied to Galactica’s predominantly scientific dataset.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.52202/075280-2191

Cites

149 of 149 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

29 Valid · 2 Not valid · 8 Not checked · 138 Own results and interpretations

89 more claims · 7 with their cited sources checked

Create a free account to see all 177 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 49
  • It may state two separate claims as one. 45
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 149 references

Research works printed but not linked: 9

  1. Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020.
  2. Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022.
  3. Kenneth Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187-197, Edinburgh, Scotland, July 2011. Association for Computational Linguistics. URL https://aclanthology.org/W11-2123.
  4. Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004.
  5. Pedro Javier Ortiz Su'arez, Laurent Romary, and Benoit Sagot. A monolingual approach to contextualized word embeddings for mid-resource languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1703-1714, Online, July 2020. Association for Computational Linguistics. URL https://www.aclweb.org/anthology/2020.acl-main.156.
  6. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
  7. Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95, 2011.
  8. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.
  9. Ge Yang, Edward Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub Pachocki, Weizhu Chen, and Jianfeng Gao. Tuning large neural networks via zero-shot hyperparameter transfer. Advances in Neural Information Processing Systems, 34:17084-17097, 2021.

Cited by

Items 201–253 of 253

Metadata

Authors
Muennighoff, N., Rush, A., Barak, B., Scao, T. L., Tazi, N., Piktus, A., Pyysalo, S., Wolf, T., & Raffel, C.
Venue
Advances in Neural Information Processing Systems 36
Published
2023
Pages
50358-50376
ISSN
1049-5258 (ISSN-L)
OpenAlex
W7133188041
Cited by
47
References
91
Citations by year
2026: 7 · 2025: 21 · 2024: 15 · 2023: 4
FWCI
8.42
Type
conference paper
Language
en
Open access
Licence
other-oa
Locations
  • neural information processing systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, other-oa)
  • UTUPub (University of Turku) (submittedVersion, other-oa)
  • IRIS Research product catalog (Sapienza University of Rome) (submittedVersion, other-oa)
  • arXiv (Cornell University)
Funders
  • National Science Foundation (2134157)
  • U.S. Department of Energy (DESC0022199)
  • European Commission (101070350)
  • China Scholarship Council
  • Defense Advanced Research Projects Agency (W911NF2010021, DE-SC0022199)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
language model scaling, training data efficiency, large language models

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-28, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.