Datasets:

---

nebius

/

SWE-rebench-V2

Like 74

Follow

Nebius 335

"fix lines readable" = lines the reference patch adds to non-test .py files (at least 20 characters, not comments, absent from base) that appear in the file at the same path in the release tag's source tree, ignoring whitespace; a file that was later moved or renamed counts as not readable. Counts use the release tag's source on GitHub, not the files inside the image. "compare" is GitHub compare from the fix PR's merge commit to the release tag.

Columns as in Appendix A, with the installation's version and location instead of the cache. pypa__packaging-29: the release history diverged from the fix, so it is not a leak. pypa__setuptools-1389: the image also has setuptools 72.1.0 in the conda base environment (and under /opt/conda/pkgs) and 75.1.0 in swebench_matterhorn; we counted only the Python 3.7 installation (68.0.0).

Every docker run uses the upstream image by digest, offline (--network none) and with --rm, and prints only paths, versions, commit hashes and counts. Platform: linux/amd64. Work dir: /<repo name>.

1. Find the project's release in pip's HTTP cache. Save this script as scan_pip_http_cache.py:

# scan_pip_http_cache.py — run inside the container: python3 - <project-name> < scan_pip_http_cache.py
import os, re, sys, tarfile, zipfile
try:
    import importlib.metadata as md
except ImportError:
    md = None
want = re.sub(r"[-_.]+", "-", sys.argv[1].lower())
DIST = re.compile(r"(?:[^/]+/)?([A-Za-z0-9_.-]+?)-([0-9][^/]*?)\.(?:dist-info|egg-info)(?:/|$)")
SDIST = re.compile(r"([A-Za-z0-9_.-]+)-([0-9][^/]*)/(?:setup\.py|pyproject\.toml)$")
roots = []
for dp, dn, _ in os.walk("/"):
    if dp.startswith(("/proc", "/sys", "/dev")) or dp.count("/") > 6:
        dn[:] = []; continue
    for d in list(dn):
        p = os.path.join(dp, d)
        if p.endswith(("/.cache/pip", "/.cache/uv")) or d == "pip-cache":
            roots.append(p); dn.remove(d)
print("cache dirs:", roots)
for root in roots:
    for dp, _, fn in os.walk(root):
        for f in fn:
            p = os.path.join(dp, f)
            try:
                head = open(p, "rb").read(2)
                names = (zipfile.ZipFile(p).namelist() if head == b"PK"
                         else tarfile.open(p).getnames() if head == b"\x1f\x8b" else [])
            except Exception:
                continue
            for nm in names:
                m = DIST.match(nm) or SDIST.match(nm)
                if m and re.sub(r"[-_.]+", "-", m.group(1).lower()) == want:
                    print("cached:", m.group(1), m.group(2), "<-", p); break
try:
    print("installed:", md.version(want) if md else "unknown")
except Exception as e:
    print("installed: none (%s)" % type(e).__name__)

# keras-team__keras-20609: expected "cached: keras 3.12.1" and "installed: 3.7.0"
docker run --rm -i --network none --user root --entrypoint python3 \
  docker.io/swerebenchv2/keras-team-keras@sha256:966e9dada6508a1aa977fbdcb78ab1ec5336330bdd26e674bbf763c54b43212c \
  - keras < scan_pip_http_cache.py

2. Check that the release contains the fix PR's merge commit (no Docker needed):

SHA=$(gh api repos/keras-team/keras/pulls/20609 --jq .merge_commit_sha)
gh api repos/keras-team/keras/compare/$SHA...v3.12.1 --jq .status   # "ahead" or "identical"

3. Count the fix lines that are readable at the same path. Save the row's patch field as fix.patch, check out the project at the row's base_commit as base/, and check out the release tag as rel/ (for example git clone --depth 1 --branch v3.12.1 https://github.com/keras-team/keras rel). Use the tag's source tree rather than an unpacked wheel: a wheel drops the src/ prefix of src-layout projects, so the same-path lookup would miss. Then run:

# count_fix_lines.py: usage: python3 count_fix_lines.py fix.patch BASE_DIR RELEASE_DIR
#   fix.patch   : the row's `patch` field saved to a file
#   BASE_DIR    : the project's source tree at the row's base_commit
#   RELEASE_DIR : the project's source tree at the release tag
# Counts the lines the patch adds to non-test .py files (>= 20 characters, not comments,
# absent from the base version of that file) and how many of them appear in the file at
# the same path in the release, ignoring whitespace. A file that was later moved or renamed
# counts as not readable. Prints counts only.
import os, re, sys

patch, base, rel = sys.argv[1:4]
norm = lambda s: re.sub(r"\s+", "", s)

def is_test(p):
    b, P = os.path.basename(p), "/" + p
    return ("/tests/" in P or "/test/" in P or "/testing/" in P
            or b.startswith("test_") or b.endswith("_test.py") or b == "conftest.py")

def read(root, path):
    p = os.path.join(root, path)
    return norm(open(p, errors="replace").read()) if os.path.exists(p) else None

files, cur = {}, None
for line in open(patch, errors="replace"):
    m = re.match(r"diff --git a/\S+ b/(\S+)", line)
    if m:
        cur = m.group(1); files.setdefault(cur, []); continue
    if cur and line.startswith("+") and not line.startswith("+++"):
        files[cur].append(line[1:])

n = k = 0
for path, added in files.items():
    if is_test(path) or not path.endswith(".py"):
        continue
    btxt = read(base, path) or ""
    rtxt = read(rel, path)
    for a in added:
        s = a.strip()
        if len(s) < 20 or s.startswith("#") or norm(s) in btxt:
            continue
        n += 1
        k += rtxt is not None and norm(s) in rtxt
print(f"fix lines readable at the same path: {k}/{n}")

python3 count_fix_lines.py fix.patch base rel   # keras-team__keras-20609: 20/29

4. Another installation of the project (python-attrs__attrs-1321): list its version and which files differ from the work tree. No file content is printed.

docker run --rm --network none --entrypoint sh \
  docker.io/swerebenchv2/python-attrs-attrs@sha256:3e949c5dd2557c542d20524c697f7e6ffd86cbf22356bef7cc9e59fd05bd6af6 -c '
  SP=/opt/conda/envs/swebench_matterhorn/lib/python3.12/site-packages
  ls -d $SP/attrs-*.dist-info
  cd /attrs && git rev-parse HEAD
  diff -rq --exclude=__pycache__ /attrs/src/attr $SP/attr'