Riazi-8B: An Urdu Large Language Model for Mathematical Reasoning

Researchers have introduced Riazi-8B, an Urdu large language model designed for mathematical reasoning, aiming to address a performance gap in low-resource languages where English-centric training has dominated recent advances [1]. The model was submitted on 24 June 2026 and is built through a two-step adaptation process: continued pre-training on Urdu Wikipedia followed by supervised fine-tuning on Urdu Chain-of-Thought data derived from the GSM8K dataset [1][2]. The developers note that Urdu lacks both reasoning-oriented resources and models adapted for multi-step mathematical problem solving, which has limited the applicability of recent progress to Urdu-speaking users [2]. Riazi-8B was evaluated on the MGSM-Urdu benchmark against existing Urdu instruction-tuned models. The results showed consistent improvements across four dimensions: answer correctness, reasoning quality, response completeness, and Urdu generation quality [1][2]. The findings indicate that combining Urdu language adaptation with reasoning-focused fine-tuning is an effective strategy for extending mathematical reasoning capabilities to low-resource languages [2]. The work addresses a documented imbalance in large language model development. Recent gains in mathematical reasoning have relied heavily on English-centric training resources and benchmarks, causing reasoning performance to degrade substantially in languages such as Urdu [2]. The Riazi-8B approach of pairing continued pre-training on a target language corpus with task-specific supervised fine-tuning mirrors transfer-learning strategies observed in other domains, where models trained on larger datasets are fine-tuned on smaller, specialized datasets to improve performance [4]. The release of Riazi-8B contributes to a broader effort to make reasoning-oriented artificial intelligence tools accessible across linguistic boundaries. The model's two-step methodology offers a replicable template for other low-resource languages that similarly lack dedicated mathematical reasoning datasets and adapted models [2].