md-nishat-008/TigerCoder-1B
TigerCoder-1B is a 1 billion parameter code generation LLM developed by Nishat Raihan, Antonios Anastasopoulos, and Marcos Zampieri from George Mason University. Built on the Gemma 3 architecture with a 32,768 token context length, it is specifically fine-tuned for generating code from Bangla prompts. This compact model excels at Python, C++, Java, and JavaScript code generation, outperforming larger general-purpose models on Bangla coding benchmarks, and is designed to run efficiently on laptop GPUs or CPUs.
Loading preview...
TigerCoder-1B: Compact Bangla Code Generation
TigerCoder-1B is a 1 billion parameter language model from the TigerCoder family, specifically designed for generating code from Bangla language prompts. Developed by Nishat Raihan, Antonios Anastasopoulos, and Marcos Zampieri at George Mason University, this model is fine-tuned from TigerLLM-1B-it using the 300K instruction-code pairs from the Bangla-Code-Instruct dataset.
Key Capabilities
- Bangla-centric Code Generation: Specializes in generating Python, C++, Java, and JavaScript code when prompted in Bangla.
- Compact and Efficient: With 1 billion parameters and a 32,768 token context length, it is optimized for deployment on resource-constrained hardware like laptop GPUs or CPUs.
- Strong Performance: Outperforms larger general-purpose models (e.g., GPT-3.5, Gemini-Flash 2.5, Gemma-3 27B) on Bangla code generation benchmarks like mHumanEval and MBPP, demonstrating that targeted data can surpass model scale for specific tasks.
- Improved Checkpoint: The current TigerCoder-1B model scores 2-4% higher than the version detailed in the LREC 2026 paper.
Good For
- Developers and researchers working on code generation in Bangla.
- Applications requiring efficient, localized code generation on edge devices.
- Scenarios where general LLMs struggle with non-English programming instructions.
- Educational tools for programming in Bangla.