arun11karthik/cellsense-fim-3b
arun11karthik/cellsense-fim-3b is a 3.1 billion parameter, long-context, fill-in-the-middle (FIM) code-completion model developed by Arun Karthik. Fine-tuned from Qwen2.5-Coder-3B, it is specifically designed for Jupyter notebooks with a 32K-token context window. This model excels at providing context-aware code completions by integrating repository files, local imports, and task-specific information, making it highly effective for in-notebook development workflows.
Loading preview...
CellSense-FIM 3B: Notebook-Native Code Completion
CellSense-FIM 3B is a specialized 3.1 billion parameter model, fine-tuned from Qwen/Qwen2.5-Coder-3B, designed for fill-in-the-middle (FIM) code completion within Jupyter notebooks. It features a substantial 32,768-token context window, enabling highly relevant and accurate suggestions.
Key Capabilities
- Notebook-Native Context Awareness: Unlike general code models, CellSense-FIM 3B processes notebooks with a deep understanding of their structure and surrounding environment.
- Repository-Aware Completions: Integrates surrounding files from the same repository into its context, ensuring completions align with project-wide helpers, constants, and conventions.
- Local-Import-Awareness: Pulls relevant source and signatures for local imports, allowing the model to complete calls to user-defined code with correct signatures.
- Task-Awareness: Prioritizes files recently read or edited, reflecting the user's current focus for more pertinent suggestions.
- Optimized for JupyterLab: Best paired with the CellSense JupyterLab Plugin, which automatically assembles the complex context required by the model, eliminating the need for manual prompt engineering.
Performance
Evaluated on a private CellSense FIM dataset, CellSense-FIM 3B demonstrates significant improvements over its base model (Qwen2.5-Coder-3B) and general models like Qwen3-4B in metrics crucial for completion quality. It achieves an edit similarity of 0.73 (up from 0.09) and a BLEU score of 53.3 (up from 4.9), alongside improved token accuracy and CodeBLEU.
Good for
- Jupyter Notebook Development: Ideal for developers working extensively in Jupyter environments who require intelligent, context-aware code completion.
- Enhanced Code Quality: Helps maintain consistency and correctness by respecting project conventions and local import structures.
- Streamlined Workflow: Reduces manual effort in code completion, especially for complex, multi-file projects within a notebook context.