Hugging Face's logo

Datasets:
CrossSWE
/
CrossSWE-Bench

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
software-engineering
cross-repository
code-repair
patch-generation
benchmark
docker
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
CrossSWE-Bench / docs /running.md
Silin-Chen's picture
Silin-Chen
CrossSWE-Bench: current 171-instance release
0e239c6
preview code
|
Raw History Blame Contribute Delete
6.56 kB

Running Docker and Tests

This release uses the current, digest-pinned environment for each of the 171 tasks. The data/images.jsonl table is a convenient index; the original image descriptors are under environments/<instance-id>/image.json and retain their original hashes.

Requirements

  • A Linux host with Docker and linux/amd64 support. - Python 3.10 or later for the shared tools; task-specific entry points may declare additional requirements. - Access to the published image references in silinchen0617/crossrepo-bench on Docker Hub. - Sufficient memory and storage for the selected task. Requirements vary substantially. For example, the Deno native evaluator checks for at least 52 GiB of host-available RAM before a run.

The dataset and Docker images are accessed through their published references. Keep credentials out of patches and checked-in scripts.

Inspect and open an environment 

python3 scripts/verify_hf_release.py
python3 scripts/crossswe.py list
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

The shell command starts an isolated container with networking disabled and no host mounts. It does not inject reference patches or evaluator tests. Repository workspace locations appear in show and in the original task metadata.

If an HTTP snapshot download has discarded executable file permissions, restore the recorded permissions before using task-specific executable scripts:

python3 scripts/restore_file_modes.py

The shared crossswe.py tool handles image inspection, pulling, and interactive access. Evaluation follows each task's own adapter and native command contract. The repository includes multiple adapter families; a shared process exit code is not a substitute for native case results.

Use python3 scripts/crossswe.py evaluator --instance-id INSTANCE to inspect the published entry-point index. Static source inspection identifies candidate-mode APIs for 57 instances. The other 114 entries point to their task-specific documentation and reference/formal scripts; they are not advertised as a validated generic candidate command. This inventory is distinct from the delivered image and recorded native test evidence for all 171 tasks.

Use the evaluation directory to find each contract and adapter. These commands inspect and filter the published entry points without starting evaluation:

python3 scripts/crossswe.py list --candidate-support detected
python3 scripts/crossswe.py list --candidate-support manual
python3 scripts/crossswe.py evaluator --instance-id runtime-deno-28946 --format text

Read a task statement with crossswe.py statement, or open statements/<instance_id>.md. Both contain the same text as the main table.

Public tests in the agent image

Public tests use the provided sources and ordinary dependencies in the image. For example, the current Deno image provides a metadata preparation helper followed by its paired-repository public test entry:

IMAGE=$(python3 scripts/crossswe.py image --instance-id runtime-deno-28946)
docker pull --platform linux/amd64 "$IMAGE"
docker run --rm --platform linux/amd64 --network=none \
  --entrypoint /bin/bash "$IMAGE" -lc \
  'python3 /opt/crossswe/prepare-public-metadata.py --output-dir /tmp/public-metadata && /opt/crossswe/run-public-tests'

That entry selects 136 provider tests and six consumer tests. Its preparation mirrors the original Cargo setup; it does not apply Gold or frozen evaluator patches. Use the corresponding environment's documentation or public-test profile for other instances. Do not substitute commands from another task or a different image version.

Evaluate a candidate patch

Candidate changes are explicit and repository-specific. Keep Gold and frozen tests outside the agent's workspace. The evaluator applies the frozen tests after the candidate production changes.

For the Deno task, prepare one patch file per modified repository. A metadata-only validation checks input locks and allowed patch paths without running Docker or tests:

python3 environments/runtime-deno-28946/v1/evaluator-support/run-evaluation.py \
  --instance-dir instances/runtime-deno-28946 \
  --world candidate \
  --candidate deno=patches/deno.patch \
  --candidate deno_config=patches/deno_config.patch \
  --validate-only

To execute the candidate, use a fresh external output directory:

IMAGE=$(python3 scripts/crossswe.py image --instance-id runtime-deno-28946)
python3 environments/runtime-deno-28946/v1/evaluator-support/run-evaluation.py \
  --instance-dir instances/runtime-deno-28946 \
  --world candidate \
  --candidate deno=patches/deno.patch \
  --candidate deno_config=patches/deno_config.patch \
  --image "$IMAGE" \
  --output-dir /tmp/crossswe-deno-candidate

Consult the selected instance's evaluator documentation for its flags, setup, source projections, resource limits, and native result format. Use its indexed candidate entry for a submitted solution; reference-endpoint controllers serve a different evaluation mode.

Optional: reconstruct an image

The published Docker image is ready to pull and use. Reconstruction is optional and is separate from running a candidate evaluator. Each image descriptor identifies its context, dependency locks, immutable parent information, or image-input materializer. The layout varies by adapter family. Read its corresponding recipe before building; some contexts depend on recorded parent descriptors or split archives.

For Deno, the self-contained scratch recipe materializes 42,330 inputs and 6,741 directory records from its pinned image, including the complete SDK and provided sources. It restores two original OCI configuration files that Docker masks during container creation. No GPU-server path or unpublished local SDK tag is required by that materializer.

Result interpretation

Preserve native test names, pass/fail counts, collection and compile diagnostics, ignored cases, and unreached selections separately. A case that never executes is not a passing test, an executed failure, or an executed skip. Some original initial states have explicitly documented task-caused API absence; those records do not waive candidate or fixed-endpoint requirements.

See Release and Validation Notes.