LuxIT: A Luxembourgish Instruction Tuning Dataset from Monolingual Seed Data
LaTeLL 2026 (Language Technologies for Low-resource Languages)
We generate Luxembourgish instruction-answer pairs from native Luxembourgish texts with DeepSeek-R1 and keep the 227,507 that pass an LLM-as-a-judge quality check. Fine-tuning 14 open models of up to 15B parameters on LuxIT raises their accuracy on Luxembourgish language exams by 5.37 percentage points on average (12 of 14 improve), and 9 of 14 also improve on five downstream NLP tasks.
@inproceedings{valline-etal-2026-luxit,
title = {{LuxIT}: A {Luxembourgish} Instruction Tuning Dataset from Monolingual Seed Data},
author = {Valline, Julian and Lothritz, Cedric and Guo, Siwen and Cabot, Jordi},
booktitle = {Proceedings of the International Conference LAnguage TEchnologies for Low-resource Languages},
month = sep,
year = {2026},
address = {Fes, Morocco},
pages = {102--120},
url = {https://acl-bg.org/proceedings/2026/LaTeLL%202026/pdf/2026.latell-1.13.pdf}
}