Search-SWE Task Data

Public input data for the Search-SWE benchmark.

The dataset provides task runtime assets such as corpora, fixed vectors, training data, and public validation inputs. File formats vary by task; use the benchmark download script to restore the expected layout. File inventories, sizes, and checksums are recorded in the asset manifests.

File Mapping

The contents of each task's public data/ directory are placed directly under tasks/<task-id>/ here. For example:

HF dataset:       tasks/task-1-2/vectors.f32
Local task:       tasks/task-1-2/data/vectors.f32
Asset manifest:   tasks/task-1-2/assets.json -> data/vectors.f32

The dataset's manifest.json and the task manifests in the code repository record the expected sizes and SHA-256 checksums. Hidden evaluation queries and labels remain in the code packages' tests/data/ directories and are not included here. Fixed model weights are downloaded separately from their original repositories.

Download

From a matching checkout of the Search-SWE code repository:

python -m pip install -r scripts/requirements-assets.txt
python scripts/download_assets.py --task all --kind data

For a single task, replace all with its ID, such as task-1-1. The code repository pins this dataset to an immutable commit in each task's assets.json. The downloader verifies file sizes and SHA-256 before installing files, and verifies existing files before skipping them.

Data sources

See SOURCES.md for upstream datasets, processing notes, and licensing information.