Published in the Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026), at EACL 2026, Rabat, Morocco.
The paper introduces Sindhi-Reasoning-Instruct, a culturally grounded Sindhi instruction dataset, and fine-tunes six LLaMA and Mistral models (1B to 24B) for deductive, inductive and causal reasoning. Native, authentic data proved decisive: Mistral-Small-24B improved human quality ratings by 141% over its base version and reached 83% on inductive reasoning tasks.
Authors: Mehak Mehak, Kamyar Zeinalipour, Pireh Soomro, Cristiano Chesi, Marco Gori, Marco Maggini.
Read the paper | Code | Models and data
All the lab's language releases are listed in Language models and datasets.