ise-uiuc/magicoder

OSS-Instruct generates low-bias instruction data by grounding LLM synthesis in open-source code references, addressing inherent dataset biases in LLM-generated training data. The approach creates diverse, realistic instructions with explicit code snippets rather than purely synthetic examples. Models are available across multiple architectures (Llama2, DeepSeek) and fine-tuning strategies, with Magicoder-S-DS-6.7B achieving 76.8% on HumanEval, outperforming GPT-3.5-turbo and Gemini Ultra.

2,086 stars. No commits in the last 6 months.