StarCoder2-15B

StarCoder2-15B is a 15 billion parameter language model trained on over 600 programming languages from The Stack v2 dataset. It leverages a Transformer decoder architecture with grouped-query and sliding window attention mechanisms, enabling it to process context windows of up to 16,384 tokens. The model was trained using the Fill-in-the-Middle objective on over 4 trillion tokens using NVIDIA's NeMo framework on NVIDIA DGX H100 systems. StarCoder2-15B excels in code generation tasks, providing code snippets based on context, though it is not an instruction model and does not work well with natural language instructions directly. It offers quantized versions using bitsandbytes for reduced memory footprint, making it accessible on various hardware configurations, from CPU to multi-GPU setups.