zed-industries/zeta

Model Overview

zed-industries/zeta is a 7.6 billion parameter model, fine-tuned from Qwen2.5-Coder-7B, specifically developed by zed-industries for edit prediction within the Zed editor. This model's primary function is to intelligently suggest and generate code modifications, enhancing the developer experience.

Key Capabilities

• Edit Prediction: Optimized for predicting and generating changes to existing code.
• Code-centric Fine-tuning: Built upon a coder-focused base model and further fine-tuned on the zed-industries/zeta dataset.
• Efficient Inference: Supports advanced vLLM features like FP8 quantization and n-gram speculative decoding for faster and more efficient operation, particularly beneficial for tasks where many tokens are already present in the prompt.

Training Details

The model was fine-tuned using the zeta dataset. Both DPO (Direct Preference Optimization) and SFT (Supervised Fine-Tuning) scripts are available for those interested in replicating or further customizing the training process.

Ideal Use Cases

• Intelligent Code Editors: Perfect for integration into IDEs and code editors requiring advanced code suggestion and modification capabilities.
• Developer Tooling: Any application that benefits from predicting and applying code edits efficiently.
• Code Generation with Context: Excels in scenarios where the model needs to generate changes based on existing code context.