Search-SWE Task Data
Fixed input assets for Search-SWE, a benchmark that evaluates coding agents on building and optimizing search systems under resource constraints. Tasks span query understanding, indexing, retrieval, ranking, and representation learning.
Homepage · Task gallery · Benchmark code · Evaluation trajectories
What is included
This repository hosts task inputs: document and code corpora, multimodal data, precomputed vectors, training examples, public validation sets, selected fixed model files, and verifier-only assets. Formats and sizes vary by task.
The benchmark code repository contains task instructions, environments, asset definitions, and grading code. Model assets can come from upstream model repositories or task-specific files hosted here; each task's asset definitions select the required sources. Evaluation traces and results are published in the separate trajectories dataset.
Some tasks use public upstream datasets directly or share assets already hosted here. The number of folders in this dataset therefore differs from the number of benchmark tasks. Use the task gallery for the current task catalog.
Get started
Run an evaluation
Use Python 3.12+ on a native Linux x86_64 host with Git, Docker, and sufficient resources for your selected tasks:
python -m pip install search-swe
search-swe run
The CLI helps you select or clone a benchmark checkout, configure your agent and credentials, and choose tasks. It downloads the selected tasks' missing assets and validates their checksums before launching evaluation. You do not need to download this entire dataset first.
See the quick start for the full workflow.
Download data without running an evaluation
From a benchmark checkout, install its dependencies and preview one task's data before downloading:
git clone https://github.com/VectorSpaceLab/Search-SWE.git
cd Search-SWE
python -m pip install -e .
python scripts/download_assets.py --task reasoning-query-rewriting --kind data --dry-run
python scripts/download_assets.py --task reasoning-query-rewriting --kind data
python scripts/download_assets.py --task reasoning-query-rewriting --kind data --verify-only
Replace the task name as needed; multiple names are accepted. Omit --kind data
to include fixed model assets. Preview sizes before selecting --task all, and
allow disk space for both the Hugging Face cache and restored files. This public
dataset does not require authentication; upstream gated models may require it.
See the asset guide for cache configuration, local restoration, and download recovery.
Repository layout and file mapping
tasks/<task-name>/ Task assets, sometimes with versioned subdirectories
provenance/ Source records, processing notes, and third-party notices
manifest.json File inventory with sizes and SHA-256 checksums
SOURCES.md Upstream sources, transformations, and usage terms
Each task's assets.json in the benchmark checkout defines the source
repository, immutable revision, filename, local destination, size, and SHA-256
for its downloaded assets. For example:
HF file: tasks/memory-constrained-dense-retrieval/vectors.f32
Local file: tasks/memory-constrained-dense-retrieval/data/vectors.f32
Mapping defined: tasks/memory-constrained-dense-retrieval/assets.json
Follow the manifest's mapping rather than inferring destinations from folder names: tasks can share files or reference versioned paths. The dataset-level manifest.json is the hosted-file inventory; it does not replace the per-task asset definitions.
Versions and reproducibility
Use the immutable revisions pinned by your benchmark checkout. Different tasks
may reference different dataset commits, and main may contain multiple asset
versions. The downloader checks both size and SHA-256 before installing files
or skipping an existing copy. Retrying can reuse the Hugging Face cache.
Removing an asset from main does not remove it from historical commits.
Use the current benchmark catalog to identify active tasks.
Evaluation access boundaries
Public availability is separate from permission to use a file during evaluation.
Some large verifier-only arrays are hosted here for download but remain forbidden
solver inputs. Other held-out queries, labels, and reference results are packaged
under the benchmark tasks' tests/data/ directories.
Follow each task's instructions and environment configuration: the agent receives its permitted inputs, while verifier inputs remain isolated from the submission. See the benchmark design for the evaluation protocol.
Sources and usage terms
There is no single license covering this collection. Original source terms apply to each corpus, code snapshot, annotation set, and other input. The benchmark's code license does not replace those terms.
Consult SOURCES.md and the provenance records for upstream revisions, attribution, transformations, and third-party notices. These records include noncommercial and share-alike restrictions, as well as unresolved upstream licensing questions; publication here does not resolve them. Retain the applicable source attributions when using or redistributing assets.
Report missing files, checksum mismatches, or documentation issues through Search-SWE issues, including the task name, benchmark revision, and asset path.
ICSI meeting memory public validation
The five public ICSI meeting-memory examples are available in tasks/icsi-meeting-memory/data/validation/ as queries.jsonl, ground_truth.jsonl and evidence.jsonl. This addition preserves the previously bundled bytes and excludes the 143 held-out examples. See provenance and applicable source terms.
Instruction-following-embedding public inputs
tasks/instruction-following-embedding/data/ adds the 1,717-record processed training set and the existing 20-group public validation set. Training data follows the upstream Apache-2.0 declaration; public evaluation annotations correspond to the MIT-labelled MTEB releases. See task provenance and its accompanying license notices. Original news corpus text and hidden groups are not included in these additions.
Khmer embedding training validation inputs
tasks/khmer-embedding-training/data/validation/ contains three public Khmer questions with English answer aliases, split into queries.jsonl and answers.jsonl. They preserve the task's existing public examples; no hidden evaluation inputs are included. These MKQA-derived records retain the Apache 2.0 source terms and attribution documented in SOURCES.md.
Incremental vector search prepared-input update (2026-11-05)
Adds 28 directly restorable corpus, public validation and verifier-only numeric files under tasks/incremental-vector-search/. Existing assets remain available. See source and derivation notes and the root manifest for per-file checksums.
- Downloads last month
- 2,786