Magicoder / OSS-Instruct

Code LLMs (ICML 2024) instruction-tuned with OSS-Instruct — 75K synthetic instruction pairs seeded from real open-source code snippets to cut the bias of purely LLM-generated data. Pre-window legacy seed from Lingming Zhang's group; the OSS-Instruct data recipe has been reused broadly across code-LLM training pipelines and fed the line that led to SelfCodeAlign.