Astro-HEP-BERT is a bidirectional transformer designed primarily to generate contextualized word embeddings for computational conceptual analysis in astrophysics and high-energy physics (HEP). Built upon Google's bert-base-uncased, the model underwent additional training for three epochs using the Astro-HEP Corpus, containing 21.84 million paragraphs found in more than 600,000 scholarly articles sourced from arXiv, all pertaining to astrophysics and/or high-energy physics. The sole training objective was Masked Language Modeling (MLM).

To optimize the model's ability to embed domain-specific language, training was conducted exclusively on entire paragraphs, rather than packing in as many sentences as possible, as often suggested in BERT tutorials. This "full-paragraphs format" preserves sentences within their original context, which is especially meaningful in academic writing where paragraphs focus on one idea.

The Astro-HEP-BERT project demonstrates the general feasibility of training a customized bidirectional transformer for computational conceptual analysis in the history, philosophy, and sociology of science as an open-source endeavor that does not require a substantial budget. Leveraging only freely available code, weights, and text inputs, the entire training process was conducted on a single MacBook Pro Laptop (M2/96GB).

For further insights into the model, the corpus, and the underlying research project (Network Epistemology in Practice) please refer to the following three papers: