Overview
StarCoder2 is the second generation of the BigCode project, an open scientific collaboration. It was trained on The Stack v2, built with Software Heritage, which covers more than 600 programming languages. The training data is documented and lets source-code authors opt out, which makes StarCoder2 a reference model for transparency.
Strengths
- Transparent, documented training data
- Good fill-in-the-middle completion
- Small sizes are efficient for autocomplete
- A popular base for fine-tuning research
Run it locally
ollama run starcoder2:7bGood fit for
Organizations that need a clear data lineage, and researchers building custom code models.
Frequently asked questions
Why is StarCoder2 called fully open?
Besides the weights, the BigCode project published the training dataset (The Stack v2), data governance tooling and a detailed technical report. Developers can also opt their code out.
Is StarCoder2 an instruction-tuned chat model?
The main releases are base models for completion. Community instruction-tuned variants such as StarCoder2-15B-Instruct exist.