Luca Soldaini

Luca Soldaini is a machine learning researcher working on scalable data methods for language models, from pretraining to post-training. As of September 2026, he is a member of the technical staff at Microsoft AI, working on MAI-Thinking models. From 2022 to early 2026, he was a lead research scientist at Ai2 and co-led Olmo, a project that released models alongside their training data, code, recipes, and checkpoints.[2]

Soldaini and Kyle Lo led Dolma, an open English corpus for language-model pretraining. Its 2024 paper describes three trillion tokens drawn from web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. Their leadership covered planning and design, infrastructure, implementation, experiments, documentation, data and compute partnerships, and discussions with legal and ethics committees. The collaborative paper received ACL 2024’s Best Resource Paper Award.[4][5]

Soldaini also developed specific parts of Dolma’s data-processing pipeline. He developed its language-filtering approach; co-developed its “quality” filter with Rodney Kinney and Zejiang Shen; and co-developed its personally identifiable information (PII) filter with Abhilasha Ravichander and Kyle Lo.[4]

Dolma pairs the released corpus with a reusable curation toolkit. The toolkit supports parallel document processing, built-in and custom taggers, and document deduplication using a Rust Bloom filter. It can run on a single machine, a cluster, or in a cloud environment, giving researchers tools to construct datasets as well as a corpus to study.[6]

Soldaini participated as an Ai2 Olmo research lead in Latent Space’s NeurIPS discussion, published on December 23, 2024 as “2024 in Open Models”.[1]

Before his Ai2 work, Soldaini studied information retrieval and retrieval-based question answering. He completed his computer science Ph.D. at Georgetown University in 2018, working with Nazli Goharian in the Information Retrieval Lab. His dissertation focused on information retrieval systems for medical experts and lay health users.[2][7]

In 2021, while affiliated with Amazon Alexa, he coauthored “Answer Generation for Retrieval-based Question Answering Systems” with Chao-Chun Hsu, Eric Lind, and Alessandro Moschitti. The work addressed a limitation of answer-sentence selection: retrieved sentences can contain correct information while still making poor standalone answers. The team trained a sequence-to-sequence transformer to generate an answer from a set of top retrieved candidates. Across three English answer-selection datasets, the authors reported improvements of up to 32 absolute percentage points in accuracy over the state of the art.[7]