arun11karthik/cellsense-fim-0.5b
arun11karthik/cellsense-fim-0.5b is a 0.5 billion parameter fill-in-the-middle (FIM) code-completion model specifically fine-tuned for Jupyter notebooks, built upon Qwen/Qwen2.5-Coder-0.5B. It supports a 32K-token context window and is optimized for repository-aware, local-import-aware, and task-aware code completion within a notebook environment. This model excels at providing contextually relevant code suggestions by considering surrounding files and user activity, making it ideal for enhancing productivity in Jupyter workflows.
Loading preview...
Overview
CellSense-FIM 0.5B is the smallest model in the CellSense-FIM family, designed for long-context, fill-in-the-middle (FIM) code completion specifically within Jupyter notebooks. Fine-tuned from Qwen/Qwen2.5-Coder-0.5B, it leverages a 32K-token context window to provide highly relevant code suggestions.
Key Capabilities
- Notebook-Native FIM: Unlike general code models, CellSense-FIM is trained on a specialized dataset that understands the unique context of Jupyter notebooks.
- Repository-Aware: It considers surrounding files from the same repository, ensuring completions align with project-wide helpers, constants, and conventions.
- Local-Import-Aware: The model pulls relevant source and signatures for sibling module imports, providing accurate completions for calls to your custom code.
- Task-Aware: Context is conditioned on files recently read and edited, reflecting the user's current focus for more pertinent suggestions.
- Enhanced Performance: Evaluation on the CellSense FIM dataset shows significant gains in completion quality metrics like edit similarity (0.11 → 0.64) and BLEU (4.7 → 43.1) compared to its base model.
Intended Use
This model is best utilized with the CellSense Jupyter Lab Plugin, which automatically assembles the necessary repository, local-import, and task context into the model's native input format, eliminating the need for manual prompt engineering. It supports deployment via vLLM for GPU-accelerated inference or Ollama for fully local, CPU-based operation, ensuring privacy as no code or context leaves the user's machine.