We Rebuilt Zeta from the Training Data Up

Zeta2 is the updated edit prediction model for Zed, offering a 30% higher acceptance rate than its predecessor, Zeta1. While maintaining its core architecture, Zeta2 introduces significant changes in context building, model training, and evaluation, enhancing its performance and user feedback integration. The model is trained on an extensive dataset of nearly 100,000 examples sourced from open-source repositories, improving its accuracy and latency. The adoption of LSP-based context retrieval allows Zeta2 to make more informed predictions by accessing surrounding code definitions. Users can explore, run, and fine-tune Zeta2's open-weight model available on Hugging Face. Zed also supports multiple edit prediction providers like Mercury Coder and Copilot Next-Edit, catering to different language and editing styles. As the development team focuses on continuous improvement, they are experimenting with Direct Preference Optimization and new prompt formats to enhance model efficiency. Users are encouraged to contribute to future advancements by enabling training data collection, and those interested in furthering the project's goals are invited to join the team.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.