hangsman/sweep-next-edit-v2-7B
The hangsman/sweep-next-edit-v2-7B is a 7.6 billion parameter Qwen2-based model fine-tuned from Qwen2.5-Coder-7B. It is specifically designed to predict the next code edit a developer will make, given the current file, recent diffs, and cursor position. This model excels at anticipating code changes and generating updated code blocks, making it highly specialized for developer tooling and code assistance. It supports a context length of 32,768 tokens, enabling comprehensive understanding of code context.
Loading preview...
Model Overview
The hangsman/sweep-next-edit-v2-7B is a 7.6 billion parameter language model, fine-tuned from the Qwen2.5-Coder-7B base model. Its core purpose is to predict the subsequent code edit a developer intends to make. This is achieved by analyzing the current file content, recent code differences (diffs), and the developer's cursor position.
Key Capabilities
- Next Edit Prediction: Predicts which code block a developer will modify next and how, generating the updated code.
- Contextual Understanding: Utilizes a 32,768-token context window, incorporating file content around the cursor, cross-file context from related files (retrieval chunks), and recent edit history.
- Specialized Prompt Format: Employs a unique prompt structure with
<|file_sep|>delimiters and a<|cursor|>marker to precisely define the input context and target output. - Fine-tuning Methodology: Developed using a multi-stage fine-tuning process: Supervised Fine-Tuning (SFT), followed by Generative Reinforcement Learning from Human Feedback (GRPO), and finally Direct Preference Optimization (DPO).
Good For
- Developer Tooling: Ideal for integration into IDEs or code editors to provide intelligent code completion, refactoring suggestions, or automated code modification.
- Code Assistance: Enhancing developer productivity by anticipating coding intentions and generating relevant code changes.
- Automated Code Generation: Generating specific code blocks based on a given context and predicted edit intent.