| --- |
| dataset_info: |
| features: |
| - name: serial_number |
| dtype: int64 |
| - name: task_id |
| dtype: string |
| - name: instruction |
| dtype: string |
| - name: image_sequence |
| dtype: string |
| - name: json_data |
| dtype: string |
| - name: num_steps |
| dtype: int64 |
| - name: num_images |
| dtype: int64 |
| - name: images |
| list: image |
| splits: |
| - name: train |
| num_bytes: 1678016377 |
| num_examples: 736 |
| - name: test |
| num_bytes: 759410992 |
| num_examples: 315 |
| download_size: 2410022437 |
| dataset_size: 2437427369 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| - split: test |
| path: data/test-* |
| task_categories: |
| - image-text-to-text |
| tags: |
| - healthcare |
| - medical |
| - gui-automation |
| - vlm-agent |
| --- |
| |
| # CareFlow Benchmark |
|
|
| [**Project Page**](https://akashghosh.github.io/Care-Pilot/) | [**Paper**](https://huggingface.co/papers/2603.24157) | [**GitHub**](https://github.com/AkashGhosh/CarePilot) |
|
|
| **CareFlow** is a high-quality human-annotated benchmark for long-horizon software workflows across medical annotation tools, DICOM viewers, EHR systems, and laboratory information systems. It was introduced as part of the paper "CarePilot: A Multi-Agent Framework for Long-Horizon Computer Task Automation in Healthcare". |
|
|
| The benchmark is designed to evaluate vision-language models (VLMs) on complex, multi-step interactions in domain-specific medical contexts. |
|
|
| ## Dataset Summary |
|
|
| CareFlow covers four major categories of clinical software: |
|
|
| | Category | Platforms | |
| |---|---| |
| | DICOM viewing & infrastructure | Orthanc, Weasis | |
| | Medical image computing & annotation | 3D Slicer | |
| | Hospital information & EMR systems | OpenEMR | |
| | Laboratory information systems | OpenHospital (OOD) | |
|
|
| ### Dataset Statistics |
|
|
| | Split | Tasks | Avg. Steps | Min | Max | Actions | |
| |---|---|---|---|---|---| |
| | Train | 735 | 12.7 | 7 | 22 | 6 | |
| | Test | 315 | 12.9 | 9 | 24 | 6 | |
| | **Total** | **1050** | — | — | — | **6** | |
|
|
| ### Action Space |
|
|
| The benchmark defines 6 primary atomic semantic actions: |
| - `CLICK`: Move the cursor and click at the specified item. |
| - `SCROLL`: Scroll the active view vertically or horizontally. |
| - `ZOOM`: Adjust the magnification level of the displayed image or view. |
| - `TEXT`: Type a string into the focused input field. |
| - `SEGMENT`: Create or edit a segmentation / ROI on the medical image. |
| - `COMPLETE`: Mark the workflow or task as finished. |
|
|
| ## Usage |
|
|
| To run the CarePilot agentic pipeline on the CareFlow dataset, navigate to the `Agentic_Pipeline` directory in the [official repository](https://github.com/AkashGhosh/CarePilot) and use the following command: |
|
|
| ```bash |
| python main.py --mode dataset --max_tasks 5 |
| ``` |
|
|
| To generate Critic-augmented trajectories (SFT Data) from the training set: |
|
|
| ```bash |
| python main.py --mode dataset --max_tasks 735 --start_task 0 |
| ``` |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{ghosh2026carepilot, |
| title={CarePilot: A Multi-Agent Framework for Long-Horizon Computer Task Automation in Healthcare}, |
| author={Akash Ghosh and Tajamul Ashraf and Rishu Kumar Singh and Numan Saeed and Sriparna Saha and Xiuying Chen and Salman Khan}, |
| booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, |
| year={2026}, |
| } |
| ``` |
