daipath/qupath-osworld-benchmark

QuPath OSWorld benchmark tasks

This manual-gated public dataset contains the complete agent-run bundles for five QuPath GUI task versions.

Directory	Task	Status
QP-001/	Focused tumor–immune review	final
QP-002/	Four-class tissue composition review	final
QP-003/	Regional TIL hotspot review	final
QP-004/	Legacy landmark-guided overlay registration	legacy
QP-004_v2/	Rendered attention-map overlay registration	recommended

QP-004 is retained for audit and comparison; QP-004_v2 is the recommended overlay-registration task.

Running a task

1. Pass <task>/instruction.txt to the GUI agent unchanged.
2. Stage only the files listed in <task>/layout.txt at the specified guest paths.
3. Apply the OSWorld budget and step unit in <task>/task_metadata.json.
4. Collect the submission paths named in the instruction.

Do not expose README, attribution, license, or manifest files to the agent unless the harness separately requires them.

Each <task>-public-trial.zip is a self-contained copy of the corresponding public task directory, including the complete WSI and all agent-visible inputs. The browsable directory and public-trial ZIP contain the same task files.

Oracle separation

This repository intentionally excludes private GT, evaluator code, official annotation masks, answer-bearing case provenance, solved outputs, screenshots, and trajectories. Those remain host-side so an agent with repository access cannot retrieve the answers.

Licensing

This is a mixed-license collection. QP-001 through QP-003 use TIGER-derived assets governed by the terms in each task's LICENSE-TIGER.txt (including CC BY-NC 4.0 for the applicable training data). QP-004 and QP-004_v2 use the CC0 OpenSlide test image. See each task's ATTRIBUTION.md.

Complete ZIP bundles are provided for all five tasks, including the three large TIGER WSI files.