Galactica

Language Revolution in Scientific Research

Scientific progress faces a major obstacle of excess information. The explosive growth of scientific materials and data has made it increasingly difficult to discover useful insights in the midst of a vast amount of information. Scientific knowledge is accessed today through research engines, but they cannot regulate scientific knowledge alone. Galactica AI aims to change that.

In this article, we address Galactica: a large language model capable of storing, integrating and producing scientific knowledge. Galactica is trained on a vast scientific collection of papers, references, knowledge bases and many other sources.

Galactica outperforms current models in a range of scientific tasks. For example, in technical knowledge tests such as LaTeX equations, Galactica exceeds the latest GPT‑3 benchmark by 68.2% versus 49.0%. It also offers strong performance on the MMLU sports test (41.3% vs 35.7%) and the MATH test (20.4% vs 8.8%). It achieves the best results so far on sub‑tasks such as PubMedQA (77.6%) and MedMCQA dev (52.9%). Although not trained on a public dataset, Galactica surpasses BLOOM and OPT‑175B on the BIG‑bench test.

We believe these findings illustrate the potential of language models as a new front for science. Therefore, we have made the model available to the scientific community.

However, challenges remain. The provider removed the public demo because large language models like Galactica can generate inaccurate and unreliable outputs, even when trained on high‑quality scientific and academic data. Public understanding of the strengths and weaknesses of these models is essential for their responsible development.

Nevertheless, Galactica represents an important step in scientific research, and we stand at the threshold of a linguistic revolution that will help scientists explore and derive knowledge in ways never possible before.