ArXiv-to-Model: A Practical Study of Scientific LM Training
This work provides practical engineering insights for researchers with moderate compute budgets aiming to build domain-specialized models, but it is incremental as it does not propose a novel architecture.
The authors tackled the under-documented process of training domain-specialized scientific language models from raw arXiv sources by conducting a detailed case study, resulting in a 1.36B-parameter model trained on 52B tokens with stable convergence and insights into preprocessing, tokenization, and infrastructure bottlenecks.
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.