CrossSWE-Bench

Software Engineering Across Repository Boundaries

Dataset · Quick Start · Evaluation · Data Format

---

CrossSWE-Bench is a benchmark of 171 cross-repository software engineering tasks drawn from real upstream projects. Tasks require coordinated changes across related repositories, such as a library and its consumer, an API and its implementation, or a build tool and its dependency.

Each instance includes its problem statement, pinned repository revisions, reference patches, test materials, evaluation contract, and a container environment identified by an immutable digest. The release covers 350 repository entries across 165 distinct upstream repositories.

Latest Release

October 30, 2026 — The complete 171-instance environment release is available. It includes all 171 pinned image descriptors, source and dependency reconstruction materials, and task-specific evaluation entry points.

Quick Start

Load the tasks

from datasets import load_dataset

tasks = load_dataset("CrossSWE/CrossSWE-Bench", "default", split="test")
print(len(tasks))  # 171
print(tasks[0]["instance_id"])
print(tasks[0]["problem_statement"])
print(tasks[0]["image_ref"])
print(tasks.features["language"].int2str(tasks[0]["language"]))
print(tasks.features["task"].int2str(tasks[0]["task"]))

Open a published image

Pull and start the selected instance's image_ref directly:

IMAGE='silinchen0617/crossrepo-bench@sha256:28db0901f554354ab43ff375ecb8b8dd5e594df5bf300da224a518c1f492fbfa'
docker pull --platform linux/amd64 "$IMAGE"
docker run --rm -it --platform linux/amd64 --network=none "$IMAGE" /bin/bash

Image reconstruction is optional. Build-input manifests and build contexts are not needed to open the published image.

Download evaluation materials

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="CrossSWE/CrossSWE-Bench",
    repo_type="dataset",
    local_dir="CrossSWE-Bench",
)

The dataset is public and does not require an HF token. Evaluation uses the selected task's test materials and adapter. A full snapshot also includes optional image-reconstruction materials; it is not required just to pull and open an image. You can read data/crossswe-bench.jsonl directly with Python's json module.

Evaluation

Start from the immutable image recorded for an instance, apply your candidate changes to its provided repositories, and use its task-specific evaluator. Reference patches and evaluator-only tests stay outside the agent workspace.

cd CrossSWE-Bench
python3 scripts/verify_hf_release.py
python3 scripts/crossswe.py show --instance-id buildpkg-poetry-10673
python3 scripts/crossswe.py statement --instance-id buildpkg-poetry-10673
python3 scripts/crossswe.py pull --instance-id buildpkg-poetry-10673
python3 scripts/crossswe.py shell --instance-id buildpkg-poetry-10673

Container images are hosted at silinchen0617/crossrepo-bench and pinned by immutable digest. Pull the selected image_ref directly.

See Running Docker and Tests for public-test commands, candidate evaluation, platform requirements, and per-instance entry points. Build and runtime requirements vary by task. The evaluation directory lists the entry point, contract and instructions for every instance, including the 57 statically detected candidate APIs and 114 instance-specific entries.

The 171 delivered environments have recorded public-test, editable-source rebuild, image audit, and formal endpoint evidence. Initial outcomes include explicitly documented task-caused failures and unexecuted cases; unexecuted cases are never counted as passes or executed failures. See Release and Validation Notes for the exact scope.

Repository Contents

data/                 Dataset tables, task index, image registry, and release metadata
statements/<id>.md    Problem statements matching the main dataset table
instances/<id>/       Task records, reference patches, and tests
environments/<id>/    Image metadata, evaluation support, and optional rebuild materials
references/           Approved reference corrections and their source preconditions
scripts/              Inspection, download, integrity, and execution utilities
docs/                 Running instructions, data format, and release notes
licenses/             Upstream license texts, attribution, and scope records
assets/               CrossSWE logo

Data Format

The main table contains one row per task. Full task artifacts remain separate files so binary fixtures, native tests, and large dependency manifests retain their exact bytes.

An empty test_patches list does not mean an instance has no tests. This field contains diff-form test patches only. Some instances instead provide standalone test files, native test materials, or recorded image-probe scripts. Find those materials through evaluation_contract_path and its test_files entries, with the exact files and hashes under instances/<instance_id>/tests/. Use the corresponding task-specific evaluation entry point; these files are preserved in their original form rather than represented as synthetic patches.

Field	Description
instance_id	Stable canonical task identifier.
original_instance_id	Original identifier retained for traceability.
language	ClassLabel with 13 programming language categories, including Multi-language.
task	ClassLabel with eight high-level task-intent categories.
problem_statement	Complete task description.
repositories	Provided repository roles, URLs, base and fixed revisions, and workspace paths.
gold_patches, test_patches	Task-published patch contents, encodings, paths, and SHA256 values.
PASS_TO_PASS_count, FAIL_TO_PASS_count	Numbers of selected targets, at the granularity recorded for each instance.
PASS_TO_PASS	JSON-encoded list of regression-test target identifiers.
FAIL_TO_PASS	JSON-encoded list of issue-resolution test target identifiers.
evaluation_support	candidate_api_detected for a statically identified candidate interface, or instance_specific for per-instance instructions.
image_ref	Current container image pinned by digest.
task_path, evaluation_contract_path, environment_path	Paths to the full task, native evaluation contract, and environment materials.

See Data Format for source-projection and reference-revision conventions. Gold and test contents are evaluator material; exclude them from an agent's task context.

The count columns count selected targets, which may be cases or script contracts. Original identifiers can contain a fail_to_pass namespace even when a particular case belongs to P2P; membership is defined by the column. P2P/F2P identifiers preserve the available evaluation granularity: retained native/evaluator case IDs are preferred; script-level contracts use their real released entry files and are explicitly marked. These lists are test selections, not execution logs or native-body counts. See test-target metadata and identifier schema for granularity, definition hashes, and initial unexecuted records.

Licensing and Attribution

The dataset has mixed license scopes. Independently authored, non-code benchmark contributions are covered by the limited CC-BY-4.0 grant in LICENSE. Upstream code, patches, tests, and derived materials retain their original terms. The release includes the upstream scope and permission records.

See Third-Party Notices, upstream license inventory, and the detailed license report.