tonychenxyz/sweeper-bench

Sweeper Bench — QA Module Prompts

200 software debugging cases. One case per row in the test split.

This revision replaces the previous schema in place. Parquet and equivalent UTF-8 JSONL are provided. The default Hugging Face configuration loads only Parquet, so the duplicate formats do not duplicate rows.

Source: local 942_v16_Clean/data/metadata.jsonl, flattened from 942_v15_Clean metadata and workflows on 2026-09-17. Prompts follow the local v15 text (Find and fix issues in ...), not the previous Hugging Face wording.

Fields

Field	Type	Meaning
case_id	string	Stable case identifier, sweeper-001 through sweeper-200.
slug	string	Short case slug (for example abantecart).
repo	string	Full repository URL. Most are on GitHub; readeck/readeck is on Codeberg.
base_commit	string	Full SHA for the starting code snapshot.
reference_commit	string	Full SHA of the reference fix.
task_scoped	string	Local L4 QA-module prompt, without a runner template.
task_specified	string	Local L1 description of the specific bug.
pr	nullable int	Recorded PR number. Null when no PR was recorded; a commit URL is not substituted.
issue	nullable int	Recorded issue number. Null when no issue was recorded.
workflows	string	One YAML document with both the core and preservation workflows under workflows: [...]. Comments are stripped; parsed values are preserved.

Each case contains exactly one core workflow and one preservation workflow in that single workflows field.

Loading

from datasets import load_dataset
import yaml

ds = load_dataset("tonychenxyz/sweeper-bench", split="test", token=True)
case = ds[0]
doc = yaml.safe_load(case["workflows"])
core, preservation = doc["workflows"]

Authentication with access to this private dataset is required. data/test.jsonl is an alternative equivalent download.

Prediction and verification boundary

Give the prediction agent task_scoped plus the approved runner instructions and the base repository. Keep task_specified, reference_commit, pr, issue, and workflows outside its accessible workspace/context; these reveal the target bug or its solution source. Workflows are evaluator inputs, not prediction prompts.

Workflow usernames/passwords are synthetic benchmark application fixtures inherited from the source workflows, not provider credentials.

Provenance and limits

The export uses 942_v16_Clean/data/metadata.jsonl. That file was built from 942_v15_Clean/data/metadata.jsonl and each case's data/workflows/<case_id>.yaml. source-manifest.json records their hashes and export validation results. Prompt strings match the local v15 source. No reference patches, prediction outputs, scores, or agent trajectories are included.

This is a data export, not a new execution/evaluation. Use the source framework for product startup and workflow interpretation. Underlying repositories and source benchmark materials retain their respective licensing terms; this export does not assert a new blanket license.