Dango: A Strictly L1-Only Large Language Model for Studying Second Language Acquisition

A research team has introduced Dango, a 1.8-billion-parameter large language model built to enable controlled studies of second language acquisition, specifically Japanese-to-English transfer, according to a paper posted to arXiv [1]. The model addresses a limitation in prior computational SLA research, which the authors say has “predominantly relied on smaller or non-decoder models” that cannot generate open-ended text and are less useful as practical simulators of second-language learners [1][2]. Dango is designed to simulate the L2 acquisition process by first acquiring a native language—Japanese—and then learning English through fine-tuning on LLM-generated lessons [1][2]. A central technical challenge the researchers identified is L2 contamination within the monolingual pretraining corpus used for the L1 phase. When scaling to 1.8B parameters, even corpora intended to be single-language often contain English fragments that give the model premature exposure to the target language [1][2]. To mitigate this, the team developed a filtering method that reduces unwanted English exposure while retaining a realistic, minimal level of contact [1][2]. The filtered model was then fine-tuned on synthetic L2-learning lessons to replicate the acquisition process. In evaluations, Dango produced human-like L2 production patterns and outperformed both unfiltered baselines and standard multilingual models [1][2]. The authors have released the model, training data, and code to support reproducible computational SLA studies and learner-facing applications [1][2]. Computational approaches to SLA have historically been constrained by model scale and architecture. Earlier work often used encoder-only or small autoregressive models, which limited the study of free-form learner output [1][2]. The shift to a 1.8B-parameter decoder-only model allows researchers to examine phenomena such as syntactic transfer, error patterns, and interlanguage development in generated text that more closely resembles human learner production [1][2]. The release comes as the field of AI-assisted language learning draws increased attention, though the Dango paper focuses strictly on research infrastructure rather than commercial deployment. By providing a controlled environment where the L1 pretraining data can be audited for L2 contamination, the model offers a testbed for hypotheses about cross-linguistic influence that are difficult to isolate in human subjects [1][2].