md-nishat-008/TigerCoder-12B
TigerCoder-12B by md-nishat-008 is a 12.2 billion parameter code-generation LLM based on the Gemma 3 architecture, specifically fine-tuned for generating code from Bangla prompts. It excels at translating Bangla programming tasks into Python, C++, Java, and JavaScript code, outperforming larger general-purpose models on Bangla-specific coding benchmarks. This model is ideal for developers and researchers focused on code generation in the Bangla language.
Loading preview...
TigerCoder-12B: Bangla Code Generation LLM
TigerCoder-12B, developed by md-nishat-008 at George Mason University, is a 12.2 billion parameter language model from the TigerCoder family, specialized in generating code from Bangla prompts. Built on the Gemma 3 architecture, it was fine-tuned using LoRA on the Bangla-Code-Instruct dataset, which comprises 300K Bangla instruction-code pairs. This model is designed to address the challenge that most code LLMs perform significantly worse with Bangla prompts compared to English.
Key Capabilities
- Bangla-Native Code Generation: Translates programming tasks described in Bangla directly into code.
- Multi-Language Support: Generates code in Python, C++, Java, and JavaScript.
- Enhanced Performance: Outperforms general-purpose models like GPT-3.5, Gemini-Flash 2.5, and Gemma-3 (27B) on Bangla-specific coding benchmarks (mHumanEval and MBPP).
- Optimized for Bangla Prompts: Specifically trained to overcome issues where machine translation of Bangla programming terms can garble meaning.
Good For
- Developers in Bangla-speaking regions: Creating code directly from Bangla instructions.
- Research in multilingual code generation: Investigating the performance of LLMs on low-resource or non-English programming contexts.
- Educational tools: Assisting learners who prefer to interact in Bangla for coding tasks.
- Applications requiring specialized Bangla code generation: Where accuracy and native language understanding are critical.