This repository is the home of StarCoder, the BigCode community’s family of open large language models for code. It holds the fine-tuning and inference code for StarCoder and StarCoderBase, 15.5B-parameter models trained on permissively licensed source code from The Stack dataset. The project has about 7,500 stars on GitHub, is released under the Apache 2.0 license, and its code was last updated in 2024.
StarCoder was a landmark in open code AI: a powerful model whose training data was collected transparently, with governance, opt-out mechanisms and license filtering handled by the BigCode community of hundreds of researchers. The models handle more than 80 programming languages, fill-in-the-middle infilling, and long 8k-token contexts, making them useful for code completion, generation and editing inside IDEs and pipelines. This repo focuses on the reproducible side of the release โ the training and fine-tuning recipes and inference scripts โ while the weights themselves are distributed through Hugging Face. Even as newer code models have appeared, StarCoder remains a reference point for how to build and release an open code LLM responsibly.
Free. The code is open source under the Apache 2.0 license and the model weights are publicly available, subject to the BigCode OpenRAIL responsible-use license.
Researchers and engineering teams interested in open code models โ whether fine-tuning StarCoder for a private codebase or studying how a major open code LLM was trained and released.