Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi

TL;DR

This paper presents LilMoo, a 0.6-billion-parameter Hindi language model trained from scratch with a transparent pipeline, outperforming larger multilingual models and addressing linguistic inequalities in NLP for low-resource languages.

Contribution

Introduces LilMoo, a fully trained Hindi language model from scratch using a high-quality dataset and optimized training recipes, challenging the reliance on multilingual foundations.

Findings

LilMoo outperforms comparable multilingual models like Qwen2.5 and Qwen3.

A high-quality Hindi corpus (GigaLekh) was constructed using heuristic and LLM-based filtering.

Designing language-specific pretraining can match larger multilingual models at small scales.

Abstract

The dominance of large multilingual foundation models has widened linguistic inequalities in Natural Language Processing (NLP), often leaving low-resource languages underrepresented. This paper introduces LilMoo, a 0.6-billion-parameter Hindi language model trained entirely from scratch to address this gap. Unlike prior Hindi models that rely on continual pretraining from opaque multilingual foundations, LilMoo is developed through a fully transparent and reproducible pipeline optimized for limited compute environments. We construct a high-quality Hindi corpus (GigaLekh) filtered through both heuristic and learned (LLM-as-a-judge) methods, complemented by bilingual augmentation with curated English data. Using this dataset, we explore various training recipes for small-scale language models. Across comprehensive evaluation suites, LilMoo consistently outperforms comparably sized multilingual baselines such as Qwen2.5-0.5B and Qwen3-0.6B, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range.