This story on HackerNoon has a decentralized backup on Sia.
Transaction ID: g4KlmFPJkILdMVtKThfH6r9yMsyc5bygyikQzTeTdsQ
Cover

The Impact of Data Size on Transformer Training: Overfitting & Loss Dynamics

Written by @reinforcement | Published on 2025/6/21
TL;DR —
Explore how training data subsets (9M, 90M tokens) influence the cross-entropy loss in Transformers, examining overfitting and the convergence behavior on test sets.

Abstract and 1 Introduction

2 Related Work

3 Model and 3.1 Associative memories

3.2 Transformer blocks

4 A New Energy Function

4.1 The layered structure

5 Cross-Entropy Loss

6 Empirical Results and 6.1 Empirical evaluation of the radius

6.2 Training GPT-2

6.3 Training Vanilla Transformers

7 Conclusion and Acknowledgments

Appendix A. Deferred Tables

Appendix B. Some Properties of the Energy Functions

Appendix C. Deferred Proofs from Section 5

Appendix D. Transformer Details: Using GPT-2 as an Example

References

6.2 Training GPT-2

Figure 4: Vanilla Transformers trained on the 2M Question-Formation dataset following the settings in (Murty et al., 2023). The training losses stabilize at a value of approximately 1, which corroborates the result presented in Proposition 4.

Authors:

(1) Xueyan Niu, Theory Laboratory, Central Research Institute, 2012 Laboratories, Huawei Technologies Co., Ltd.;

(2) Bo Bai baibo (8@huawei.com);

(3) Lei Deng (deng.lei2@huawei.com);

(4) Wei Han (harvey.hanwei@huawei.com).

This paper is available on arxiv under CC BY-NC-ND 4.0 DEED license.

[story continues]

Written by
@reinforcement
Leading research and publication in advancing reinforcement machine learning, shaping intelligent systems & automation.
Topics and
tags
transformer-models|associative-memory|hopfield-networks|model-generalization|attention-mechanism|cross-entropy-loss|model-scaling|neural-network-performance
This story on HackerNoon has a decentralized backup on Sia.
Transaction ID: g4KlmFPJkILdMVtKThfH6r9yMsyc5bygyikQzTeTdsQ