SWE-Llama are variants of the CodeLlama model fine-tuned on software engineering tasks extracted from real-world GitHub issues and pull requests. They were introduced and evaluated on the SWE-bench benchmark.