arun11karthik/cellsense-fim-7b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The arun11karthik/cellsense-fim-7b model is a 7.6 billion parameter, long-context, fill-in-the-middle (FIM) code-completion model specifically fine-tuned for Jupyter notebooks. Built upon Qwen/Qwen2.5-Coder-7B, it supports a 32K-token context window and is uniquely trained to be repository-aware, local-import-aware, and task-aware. This model excels at providing highly relevant code completions within the complex context of a Jupyter project, significantly improving completion quality metrics like edit similarity and BLEU.

Loading preview...

CellSense-FIM 7B: Notebook-Native Code Completion

CellSense-FIM 7B is the largest model in the CellSense-FIM family, a series of 7.6 billion parameter, long-context, fill-in-the-middle (FIM) code-completion models. Fine-tuned from Qwen/Qwen2.5-Coder-7B on a private CellSense FIM dataset, it is specifically designed for Jupyter notebooks and supports a 32K-token context window.

Key Differentiators & Capabilities

  • Notebook-Native Context: Unlike general code models, CellSense-FIM is trained on a detailed context that matters for Jupyter cells.
  • Repository-Aware: Understands and utilizes surrounding files from the same repository, respecting project-wide helpers, constants, and conventions.
  • Local-Import-Aware: Integrates relevant source and signatures from sibling modules when a notebook imports local code, ensuring accurate completions for custom functions.
  • Task-Aware: Conditions completions on files recently read or edited, reflecting the user's current focus.
  • High Completion Quality: Achieves significant gains in real completion quality metrics, with edit similarity improving from 0.07 to 0.74 and BLEU from 4.9 to 57.0 compared to its base model.

Recommended Usage

This model is best paired with the CellSense Jupyter Lab Plugin, which automatically assembles the repository, local-import, and task context into the model's native training format. It can be served efficiently using vLLM for remote or local GPU inference, or with Ollama for fully local, no-GPU-required inference using GGUF quantizations.

Finetunes
2 models