MagiCoder

Open-source code LLM using OSS-Instruct synthetic data generation

Code & DevelopmentFree

MagiCoder is a family of open-source code generation models from the University of Illinois that achieves strong coding benchmarks through a novel training technique called OSS-Instruct. OSS-Instruct uses open-source code snippets as seeds to generate diverse, high-quality synthetic coding problems, bypassing the need for human-curated instruction datasets. MagiCoder-S-CL-7B matches or exceeds much larger models on HumanEval and MBPP benchmarks.

Key Features

• ✓OSS-Instruct training
• ✓7B parameters
• ✓HumanEval competitive
• ✓Open source
• ✓Synthetic data
• ✓Research model