SWEeper-Bench

SWEeper-Bench is a benchmark of 200 real bugs in 200 interactive web applications. Each task asks a coding agent to find and fix problems in a named product area. The bug shows up only after a sequence of ordinary user actions, so the agent has to explore the running application, not only the source. A repair passes when an agentic verifier, using the browser and judging only what a user would see, accepts both behavior tests for that task.

The 200 applications cover content and knowledge, IT and developer tools, collaboration, files and media, finance and business, personal tools, and commerce. Each application contributes one task. The starting point is the parent of a merged fix: base_commit is the buggy snapshot, and reference_commit is the reference repair. The agent does not receive the original issue, pull request, reference patch, repository history, or behavior tests.

What the agent is asked

task_scoped is the open-ended request given to the agent. It names a product area and does not say what is broken, where the code lives, or how to reproduce it. For example: find and fix issues in a product’s board sharing and membership.

task_specified is a detailed description of the target bug. It is held back from the agent. It exists so a run can be repeated with the bug already identified, which separates discovery from repair.

How a repair is judged

workflows is one YAML document with exactly two behavior tests:

• target reproduces the interactions that expose the bug and checks that the intended behavior is back.
• preservation checks that nearby behavior still works.

A task passes only when both tests pass. Each test has an id, a category (target or preservation), a purpose, and ordered steps. The steps are user-level instructions with explicit pass and fail criteria. The verifier follows them in the browser on a fresh copy of the application with the agent’s patch applied. Setup failures, such as a login that never succeeds, are not product verdicts.

Usernames and passwords inside the steps are synthetic fixtures for the benchmark applications.

Fields

Field	Type	Meaning
case_id	string	sweeper-001 through sweeper-200
slug	string	Short name of the application
repo	string	Repository URL
base_commit	string	Full SHA of the buggy snapshot
reference_commit	string	Full SHA of the reference fix
task_scoped	string	Open-ended prompt for the agent
task_specified	string	Detailed description of the target bug. Not an agent input.
pr	int or null	Recorded pull request number. Null when none was recorded.
issue	int or null	Recorded issue number. Null when none was recorded.
workflows	string	YAML with one target test and one preservation test

Loading

from datasets import load_dataset

sweeper = load_dataset("EVIGBYEN/SWEeper-Bench", revision="main", split="test")
task = sweeper[0]

What to give the agent

Give the prediction agent task_scoped, the repository at base_commit, and the running application. Keep task_specified, reference_commit, pr, issue, and workflows out of its workspace. Those fields name the target bug or point at the fix. The behavior tests are verifier inputs, not prediction prompts.

This dataset does not include product images, reference patches, model patches, scores, or agent trajectories. The source repositories keep their own licenses.

Citation

@misc{sweeperbench2026,
  title        = {SWEeper-Bench: Can Agents Discover Bugs in Interactive Software?},
  author       = {Yang Yao and Haozhe Chen and Bingyi Kang and Karthik R Narasimhan and Zhuang Liu},
  year         = {2026},
  note         = {Preprint. Yang Yao and Haozhe Chen contributed equally.}
}