Lucas Bandarkar

I'm a third-year A.I. Ph.D. student in the Computer Science department at UCLA. I'm advised by Nanyun (Violet) Peng in the PLUS Lab and study multilingual NLP. In particular, I am interested in the interpretability of cross-lingual representations and modularity in LLMs. I am supported by the Amazon AI PhD Fellowship.

I did my undergrad at UC Berkeley, and worked in Marti Hearst's NLP lab under the mentorship of Philippe Laban. I then spent over two years as a research data scientist at Meta/Facebook AI working on large-scale multilingual NLP systems, generally focusing on model evaluation, data resources, and global language strategy for a suite of production models such as machine translation, language identification, and text embeddings. Notably, I led the development of the Belebele dataset. 

During my Ph.D., I have completed research internships back at Meta in the multilingual GenAI team and at Google Research Australia on Gemini multilinguality.

• model modularity: model merging, mixture-of-experts, implicit modularity, adapters

• multi-/cross-lingual representations: cross-lingual transfer, "language-agnostic" representations, tokenization, model interpretability

• multilingual data & evaluation: language identification, data annotation & resource creation, embeddings evaluation, translation evaluation

Applications: multilingual embeddings, LLM language adaptation, LMs in low-resource languages, language identification, machine translation

(in progress) Ph.D. in Computer Science, UCLA

Sep 2023 - current

B.A. in Statistics, Data Science, UC Berkeley

Aug 2017 - May 2021