Restore BugsInPy historical static repairs with verified cached images

bugsinpy-original_environment-reliable/README.md

@@ -1,42 +1,41 @@
 # BugsInPy
 
-The pinned [native BugsInPy release](https://github.com/soarsmu/BugsInPy/tree/11c5f1eea954a42132cfd06bf257766a7963e0fd) lists **501 bugs**. We converted **500**; Keras 12 has identical buggy and fixed commits and produced no task.
+## Dataset stats
 
-TaskTrove used GPT-4o-mini to turn bug-fix diffs into self-contained Python tasks, with generated tests and instructions. We replaced those generated tasks with the actual project repositories at their buggy commits, using the upstream tests and reference fixes recorded by BugsInPy.
+- **Tasks:** 501 native bugs = **489 retained (all 489 converted/patched)** + **11 archived** + **1 not converted** (Keras 12 has identical buggy/fixed commits). Includes 462 retained instruction rewrites.
+- **Images:** native source supplied none; conversion introduces **11 shared image definitions**. **11 verified cached Apptainer bundles** cover all 489 retained tasks, pinned by checksum and immutable image commit in `tasks.parquet`.
 
-- **Docker images:** the native benchmark supplied no Docker images. Our retained tasks use **11 shared image definitions**; this was a new conversion, not image consolidation.
-- **Tasks: 501 original bugs; 489 kept (all converted), 11 excluded and 1 not converted.** The 462 reviewed instruction changes remain included.
+## Dataset characteristics
 
-## Stage 3: Build/setup and environment compatibility
+Real Python repositories at buggy commits, upstream regression tests and reference fixes from [BugsInPy](https://github.com/soarsmu/BugsInPy/tree/11c5f1eea954a42132cfd06bf257766a7963e0fd), replacing TaskTrove's synthetic code/tests.
 
-- **489 retained tasks converted:** packaged the buggy repositories and pinned runtime and test dependencies for reproducible setup.
-- **247 tasks:** added compiler tools or system development libraries to their images.
-- **1 task excluded:** YouTube-DL 40 concerns an older-Python bug that its recorded Python 3.7 environment cannot exercise.
+<details><summary>Domains</summary>
 
-## Stage 4: Reference solution reward 1
+**Debugging; Software Engineering.** Repair crashes, incorrect results and missing behavior in projects including pandas, Black, Luigi, spaCy, Tornado and youtube-dl.
 
-- **489 tasks adapted:** installed pinned test dependencies in a clean verifier environment so existing agent-side packages could not cause installation to be skipped.
-- **66 tasks:** refreshed restored source-file timestamps so Python would not reuse stale compiled bytecode after applying the fix.
-- **169 pandas tasks:** refreshed fixed source timestamps before rebuilding compiled extensions.
-- **1 task excluded:** HTTPie 4 requires the live httpbin.org service, so a correct fix can fail when the service or network is unavailable.
+</details>
 
-## Stage 5: No-op reward 0
+<details><summary>Capabilities</summary>
 
-**6 tasks patched:**
+**Exploration; Error Recovery.** Find relevant code in unfamiliar repositories, reproduce failures and iterate against regression tests. Labels follow [CLI-Universe](https://arxiv.org/abs/2606.22883v1).
 
-- **1 Ansible task:** added an existing upstream regression that detects the CPU-counting bug; the original selected test passed without the fix.
-- **4 tasks (Scrapy 33 and YouTube-DL 38, 39 and 42):** moved missing-function lookups into their tests so unfinished code fails a test instead of stopping test collection.
-- **1 YouTube-DL task:** added offline chapter-extraction checks covering JSON chapter data and the description fallback.
+</details>
 
-**3 tasks excluded:** Keras 15 and Scrapy 7 and 11 passed their selected tests without the reference fix.
+**Intended benchmark transfer:** SWE-bench Verified random-100; DeepSWE 1.1, through repository-level issue diagnosis and test-driven repairs.
 
-## Task-quality review retained
+## Per-stage breakdown
 
-- **462 reviewed instruction changes preserved.** The automatically added timeout/anti-cheating suffix is removed from all 489 retained tasks.
-- **6 existing exclusions preserved:** Ansible 12, Luigi 8, TheFuck 3 and 16, and Tornado 6 required implementation details that could reject alternative correct fixes. Black 23's selected test did not cover the issue described in its instruction.
+Counts overlap. [Historical evidence](https://huggingface.co/datasets/FWeindel/validated-tasks/blob/579dbfd3edf67734e7725129c62dcf5975945976/runs/2026-10-08-bugsinpy-original-fc23c4eacb04.json): stages 3/4/5 passed before instruction review; stage 1 reran afterwards.
 
-## Images
+- **1, rule-based/static:** 489 retain the timeout/anti-cheating suffix required by `check-instruction-suffix.sh` (427 carried the historical `anti-cheat-instruction` change label). **0 static archives**; 7 recorded `warning:relative-instruction-path` findings remained warnings. The contract excluded 13 checks. Separate instruction review preserved **462 rewrites** and archived **6** for tests requiring specific implementations (5) or testing a different issue (1), recorded as `flagged_by_instruction_loop`, not static-check failures.
+- **3, build/setup:** **489 converted** to shared Apptainer-compatible environments with pinned dependencies; **247** include compiler/system libraries. Shared Marin/Harbor bridge fixes addressed container startup, uploads and certificate mounts. **1 archived** (`environment_task_mismatch`): YouTube-DL 40 requires older-Python behavior absent from its recorded environment.
+- **4, reference reward 1:** **489 passed**. Clean verifier dependency installs for **489**; source-timestamp fixes for **66**, plus **169 pandas** native rebuilds. **1 archived** (`flagged_by_instruction_loop`): HTTPie 4 depends on live httpbin.org availability.
+- **5, no-op reward 0:** **489 passed**. **6 verifier repairs:** Ansible CPU regression (1), missing-function assertions moved inside tests (4), offline YouTube-DL chapter checks (1). **3 archived** (`buggy-version-passes-verifier`): Keras 15 and Scrapy 7/11 passed without a fix.
 
-Dockerfiles are included in every task. **11 verified cached image bundles** are available in [cached_images](https://huggingface.co/datasets/FWeindel/validated-tasks/tree/main/bugsinpy-original_environment-reliable/cached_images). Task Parquets pin the image files and checksums to an immutable commit.
+All 11 archives were recorded as stage-0 conversion/review exclusions; the bullets group their reasons by relevant stage.
 
-[Detailed provenance and conversion notes](https://github.com/MLI-lab/OT-next-data/blob/33c7a1648a5eca45310eb4515f11e14f363fabb2/data/tasktrove_bugsinpy/README.md).
+## Resources and timing
+
+Historical final static run: Helma, x86_64, **48 CPUs / 128G**, static concurrency **24**; Python **3.12.14**, Harbor **0.8.1**, Apptainer **1.5.4-1.el9**. Harbor-validation commit `7faf878c14b6`.
+
+**Validation:** stage 1 wall time **7.3 s / 489 tasks**; per-task median **0.3 s**. **Not recorded:** mean task time, aggregate stages 3/4/5 time, total/per-task patching time, separate Marin version.

bugsinpy-original_environment-reliable/archive.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1d6b44e000602924783405674cce6b90e0dd5a915f8415df55ba610eed37ddaf
-size 30872916
+oid sha256:f448a7a911269a0a530058aa9f11cd9aa1a335958fe6d2ec5a439bc1a378aebe
+size 31154810

bugsinpy-original_environment-reliable/patch.py

@@ -2374,8 +2374,8 @@ def write_project(output: Path, rows: list[dict], manifest: list[dict], errors:
     (output / "issues.json").write_text(json.dumps(cache, indent=2, sort_keys=True) + "\n")
 
 
-def remove_publication_suffix(blob: bytes, expected_sha256: str | None) -> tuple[bytes, dict[str, bytes]]:
-    """Remove only the suffix; optionally require an exact historical archive hash."""
+def remove_publication_suffix(blob: bytes, expected_sha256: str | None, *, repack: bool = True) -> tuple[bytes, dict[str, bytes]]:
+    """Remove the suffix from files; repack=False keeps the original archive bytes."""
     files = {}
     with tarfile.open(fileobj=io.BytesIO(blob)) as archive:
         for member in archive:
@@ -2394,7 +2394,9 @@ def remove_publication_suffix(blob: bytes, expected_sha256: str | None) -> tuple
     if count != 1:
         raise ValueError('Expected exactly one automatic instruction suffix')
     files['instruction.md'] = instruction
-    restored = task_tar(files)
+    if expected_sha256 is not None and not repack:
+        raise ValueError("Assembly comparison requires repacking")
+    restored = task_tar(files) if repack else blob
     if expected_sha256 is not None and sha256(restored) != expected_sha256:
         raise ValueError('Restored task differs from the accepted instruction-review payload')
     return restored, files
@@ -2416,7 +2418,8 @@ def environment_exclusion_group(task: str) -> str:
 
 
 def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_record: Path,
-                        publication_revision: str, output: Path, payload_shards: Path | None = None) -> dict:
+                        publication_revision: str, output: Path, payload_shards: Path | None = None,
+                        include_static_repairs: bool = False) -> dict:
     import pyarrow as pa
     import pyarrow.parquet as pq
     from collections import Counter
@@ -2457,12 +2460,19 @@ def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_r
         'archived': len(archive_rows), 'exclusions_by_group': dict(Counter(x['group'] for x in exclusions)),
         'reviewed_instruction_changes_preserved': len(reviewed),
         'quality_exclusions_preserved': sum(x['group'] == 'task-quality-review' for x in exclusions),
-        'automatic_suffix_removed': len(expected), 'source_docker_images': None,
+        'automatic_suffix_removed': 0 if include_static_repairs else len(expected),
+        'automatic_suffix_preserved': len(expected) if include_static_repairs else 0,
+        'source_docker_images': None,
         'image_consolidation': False, 'published_cached_bundles': 0,
         'historical_run': run_url, 'publication_revision': publication_revision,
         'fresh_validation': False,
         'evidence_transfer': 'Historical runtime evidence disclosed by the publication. Only the automatic instruction suffix is removed; all runtime files and file modes are preserved. Reviewed instructions and the existing quality exclusions are intentionally retained. No new runtime or teacher run.',
     }
+    if include_static_repairs:
+        summary['evidence_transfer'] = (
+            'Historical final stage-1 checks and pre-instruction-review runtime stages 3, 4 and 5. '
+            'Exact historical published task payloads retain reviewed instructions and rule-based suffix repairs. '
+            'No fresh runtime or teacher run.')
     published = pq.ParquetFile(tasks)
     schema = published.schema_arrow
     def input_rows():
@@ -2478,16 +2488,20 @@ def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_r
         if not shards:
             raise ValueError('No payload shards')
         for shard in shards:
-            for batch in pq.ParquetFile(shard).iter_batches(batch_size=8):
-                for payload in batch.to_pylist():
+            parquet = pq.ParquetFile(shard)
+            # Existing shards use one bounded task per row group. Reading across
+            # groups through iter_batches is pathological for large binary cells.
+            for group in range(parquet.num_row_groups):
+                for payload in parquet.read_row_group(group).to_pylist():
                     if payload['path'] not in annotations:
                         raise ValueError('Unexpected task in payload shards')
                     yield {**annotations[payload['path']], 'task_binary': payload['task_binary']}
     output.mkdir(parents=True)
     retained, images, seen, features = [], {}, set(), Counter()
     # Write a bounded-memory intermediate; add the measured summary metadata below.
+    # Payloads are already compressed tar archives; avoid expensive recompression.
     temporary = output/'tasks.data.parquet'
-    with pq.ParquetWriter(temporary, schema, compression='zstd') as writer:
+    with pq.ParquetWriter(temporary, schema, compression='NONE') as writer:
         for row in input_rows():
             task = row['path']
             if task in seen or task not in expected:
@@ -2499,9 +2513,15 @@ def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_r
             # Historical tar packaging varies across the review/retry runs.
             # Bind every input to its published binary hash, change only the
             # canonical suffix, and separately report exact assembly matches.
-            blob, files = remove_publication_suffix(row['task_binary'], None)
-            assembly_match = sha256(blob) == expected[task]['output_sha256']
-            features['accepted_assembly_archive_matches'] += assembly_match
+            blob, files = remove_publication_suffix(row['task_binary'], None, repack=not include_static_repairs)
+            assembly_match = None if include_static_repairs else sha256(blob) == expected[task]['output_sha256']
+            if assembly_match is not None:
+                features['accepted_assembly_archive_matches'] += assembly_match
+            if include_static_repairs:
+                # Restore the exact historical payload, including its tar metadata.
+                blob = row['task_binary']
+                with tarfile.open(fileobj=io.BytesIO(blob)) as original:
+                    files['instruction.md'] = original.extractfile('instruction.md').read()
             image = sha256(files['environment/Dockerfile'])
             images.setdefault(image, {'dockerfile': files['environment/Dockerfile'].decode(), 'tasks': []})['tasks'].append(task)
             features['isolated_verifier'] += b'env -u PYTHONPATH /tests/.venv/bin/python -m pip install' in files['tests/test.sh']
@@ -2509,8 +2529,9 @@ def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_r
             features['reference_native_rebuild_refresh'] += b'os.utime(path, None)' in files['solution/solve.sh']
             features['native_build_tools'] += b'apt-get install' in files['environment/Dockerfile']
             features['setup_and_verifier_locks'] += bool(files['setup_files/requirements.lock'].strip() and files['tests/requirements.lock'].strip())
-            row.update(task_binary=blob, content_sha256=files_digest(files.items()), stages_passed='3,4,5', run=run_url,
-                       patch_labels=[label for label in row['patch_labels'] if label != 'anti-cheat-instruction'])
+            row.update(task_binary=blob, content_sha256=files_digest(files.items()), stages_passed='1,3,4,5' if include_static_repairs else '3,4,5', run=run_url,
+                       patch_labels=row['patch_labels'] if include_static_repairs else
+                           [label for label in row['patch_labels'] if label != 'anti-cheat-instruction'])
             writer.write_table(pa.Table.from_pylist([row], schema=schema))
             retained.append({'task_id': task, 'historical_binary_sha256': before_binary,
                              'output_binary_sha256': sha256(blob), 'content_sha256': row['content_sha256'],
@@ -2523,9 +2544,10 @@ def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_r
     summary.update(shared_image_definitions=len(images), retained_runtime_features=dict(features))
     metadata = dict(schema.metadata or {})
     metadata[b'ot.environment_reliable.v1'] = json.dumps(summary).encode()
-    with pq.ParquetWriter(output/'tasks.parquet', schema.with_metadata(metadata), compression='zstd') as writer:
-        for batch in pq.ParquetFile(temporary).iter_batches(batch_size=1):
-            writer.write_table(pa.Table.from_batches([batch]).replace_schema_metadata(metadata))
+    with pq.ParquetWriter(output/'tasks.parquet', schema.with_metadata(metadata), compression='NONE') as writer:
+        parquet = pq.ParquetFile(temporary)
+        for group in range(parquet.num_row_groups):
+            writer.write_table(parquet.read_row_group(group).replace_schema_metadata(metadata))
     temporary.unlink()
     for row in archive_rows:
         row['run'] = run_url
@@ -2533,7 +2555,7 @@ def prepare_environment(tasks: Path, archive: Path, run_record: Path, assembly_r
             'https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/',
             'https://github.com/MLI-lab/OT-next-data/tree/33c7a1648a5eca45310eb4515f11e14f363fabb2/data/tasktrove_bugsinpy/')
     pq.write_table(pa.Table.from_pylist(archive_rows, schema=excluded.schema).replace_schema_metadata(metadata),
-                   output/'archive.parquet', compression='zstd')
+                   output/'archive.parquet', compression='NONE')
     (output/'summary.json').write_text(json.dumps(summary, indent=2)+'\n')
     (output/'audit.json').write_text(json.dumps({'summary': summary, 'retained': retained, 'exclusions': exclusions}, indent=2)+'\n')
     (output/'image-definitions.json').write_text(json.dumps(images, indent=2)+'\n')
@@ -2554,7 +2576,9 @@ def main() -> None:
                         help="Shared image by Python version; fields: major, minor, version")
     parser.add_argument("--render-dir", type=Path, help="Optional folder for inspectable Harbor tasks")
     parser.add_argument("--issue-cache", type=Path, help="Optional saved issue responses to reuse")
-    parser.add_argument('--validated-tasks', type=Path, help='Prepare the reviewed publication without its automatic suffix')
+    parser.add_argument('--validated-tasks', type=Path, help='Prepare the reviewed historical publication')
+    parser.add_argument('--include-static-repairs', action='store_true',
+                        help='Preserve exact historical payloads with reviewed instructions and rule-based suffix repairs')
     parser.add_argument('--validated-archive', type=Path)
     parser.add_argument('--run-record', type=Path)
     parser.add_argument('--assembly-record', type=Path)
@@ -2567,9 +2591,10 @@ def main() -> None:
         if any((args.start_at_project, args.stop_after_project, args.render_dir, args.issue_cache, args.metadata_revision)):
             parser.error('Publication preparation cannot be combined with native conversion options')
         print(json.dumps(prepare_environment(args.validated_tasks, args.validated_archive, args.run_record,
-            args.assembly_record, args.publication_revision, args.output, args.payload_shards), indent=2))
+            args.assembly_record, args.publication_revision, args.output, args.payload_shards,
+            include_static_repairs=args.include_static_repairs), indent=2))
         return
-    if any((args.validated_archive, args.run_record, args.assembly_record, args.publication_revision, args.payload_shards)):
+    if any((args.validated_archive, args.run_record, args.assembly_record, args.publication_revision, args.payload_shards, args.include_static_repairs)):
         parser.error('Publication inputs require --validated-tasks')
     try:
         import pyarrow  # noqa: F401 - fail before writing any rendered tasks

bugsinpy-original_environment-reliable/tasks.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0c6a13b40376dec7bb27fe07518d63a8d4fd720fc53ddc554b3275f693d3c1a1
-size 4377463814
+oid sha256:0c9b90adbe2e2d5ac8f56165ebfaa5e3dbde370f51b453884be5b3c0b2944c54
+size 4343661074