TigerCoder: A Novel Suite of LLMs for Code Generation in Bangla

2025-09-10

AI Summary

To address the severe scarcity of high-quality data and dedicated models for code generation in Bangla—the world’s fifth most spoken language—this work introduces MBPP-Bangla, the first Bangla code instruction dataset and evaluation benchmark. We further propose TigerCoder, the first open-source large language model (LLM) family specifically designed for Bangla code generation, with 1B and 9B parameter variants. Leveraging instruction tuning and domain adaptation, TigerCoder is trained exclusively on high-quality Bangla code data. Experimental results demonstrate that TigerCoder achieves a 11–18% improvement in Pass@1 over existing multilingual and general-purpose Bangla LMs, empirically validating the efficacy of pairing compact models with high-fidelity domain-specific data for low-resource language code generation. All components—including the dataset, models, and evaluation framework—are publicly released to foster community advancement.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Despite being the 5th most spoken language, Bangla remains underrepresented in Large Language Models (LLMs), particularly for code generation. This primarily stems from the scarcity of high-quality data to pre-train and/or finetune such models. Hence, we introduce the first dedicated family of Code LLMs for Bangla (1B & 9B). We offer three major contributions: (1) a comprehensive Bangla code instruction datasets for programming domain adaptation; (2) MBPP-Bangla, an evaluation benchmark for Bangla code generation; and (3) the TigerCoder-family of Code LLMs, achieving significant ~11-18% performance gains at Pass@1 over existing multilingual and general-purpose Bangla LLMs. Our findings show that curated, high-quality datasets can overcome limitations of smaller models for low-resource languages. We open-source all resources to advance further Bangla LLM research.

Problem

• Addresses Bangla's underrepresentation in code generation LLMs
• Overcomes scarcity of high-quality Bangla programming data
• Introduces specialized models and benchmarks for Bangla coding

Innovation

• Bangla code instruction datasets creation
• MBPP-Bangla evaluation benchmark development
• TigerCoder LLMs with 11-18% performance gains