We Rebuilt Zeta from the Training Data Up
Zeta2 is our updated edit prediction model. Its acceptance rate is 30% better than Zeta1, and it is now the default Zeta model for all Zed users.
The core architecture is the same one we described when we first launched edit predictions . What changed is everything around it: how we build context, how we train, how we evaluate models before deployment, and how we incorporate your feedback once it's deployed.
Zeta1 was hand-crafted. So when we built Zeta2, we had to also create the infrastructure necessary to craft at scale: to build not just one model, but many models so we can constantly learn and improve.
The biggest improvements came from the training pipeline. Zeta1 was trained on a hand-curated set of around 500 examples. Zeta2 is trained on nearly 100,000 examples collected on an opt-in basis in open-source licensed repositories from Zed users. These final training examples are the result of an entirely new data pipeline that collects, processes, and orchestrates training and evaluation against the collected samples. We'll talk more about how this pipeline works in an upcoming blog post. Until then, some of it is open-source that you can check out here . Latency improved as a byproduct of our efforts, so you'll experience faster predictions as well.
Predictions are only as good as their context, and if a model only sees the code immediately around your cursor, it has to guess at changes dependent on types defined elsewhere or functions from other modules. Earlier this year we shipped LSP-based context retrieval (the same infrastructure that powers go-to-definition) as part of our 0.222.2 release . With this context Zeta2 sees the types and definitions of symbols around your cursor, so it doesn't have to make guesses. If you've been using edit predictions since late February, we hope you've felt the difference.
Zeta2 is trained entirely on open source code and the latest iteration we've trained is now open-weight. It's trained on data from users who actively chose to share it, not from users who had to know to opt out. You can inspect the model, run it on your own infrastructure, or fine-tune it on your own codebase. Zeta1's training data was open source as well as its weights. That training set was small and hand-curated, drawn from users who had explicitly given consent to share their data. We're still training from open source repos, but we're not releasing the data at this scale. Users opted into collection and training, not into public data sharing.