ArXiv-to-Model: A Practical Study of Scientific LM Training

TL;DR

This paper provides a comprehensive case study on training a 1.36B scientific language model from raw arXiv LaTeX sources, detailing the pipeline, challenges, and insights for researchers with limited compute resources.

Contribution

It offers an end-to-end, transparent account of training a scientific language model from raw data, highlighting practical engineering considerations and bottlenecks.

Findings

Preprocessing choices greatly influence usable token volume.

Tokenization affects symbolic stability in training.

Storage and I/O constraints can rival compute as bottlenecks.

Abstract

While frontier large language models demonstrate strong reasoning and mathematical capabilities, the practical process of training domain-specialized scientific language models from raw sources remains under-documented. In this work, we present a detailed case study of training a 1.36B-parameter scientific language model directly from raw arXiv LaTeX sources spanning mathematics, computer science, and theoretical physics. We describe an end-to-end pipeline covering metadata filtering, archive validation, LaTeX extraction, text normalization, domain-aware tokenization, and dense transformer training under constrained compute (2xA100 GPUs). Through 24 experimental runs, we analyze training stability, scaling behavior, data yield losses, and infrastructure bottlenecks. Our findings highlight how preprocessing decisions significantly affect usable token volume, how tokenization impacts symbolic stability, and how storage and I/O constraints can rival compute as limiting factors. We further analyze convergence dynamics and show stable training behavior in a data-rich regime (52B pretraining tokens). Rather than proposing a novel architecture, this work provides an engineering-grounded, transparent account of training a small scientific language model from scratch. We hope these insights support researchers operating under moderate compute budgets who seek to build domain-specialized models.