Validation run 2026-10-08-bugsinpy-original-fc23c4eacb04
#7
by FWeindel - opened
bugsinpy-original
Keeps 489 tasks and archives 11 from BugsInPy at commit 11c5f1eea954a42132cfd06bf257766a7963e0fd.
| Archive label | Tasks | Explanation |
|---|---|---|
| patch-drop:buggy-version-passes-verifier | 3 | The unchanged buggy project already passes the selected verifier, so the test does not distinguish a repair. The unchanged buggy project already passes the selected verifier, so the test does not distinguish a repair. The reported bug concerns Windows behavior; this task runs on Linux. |
| patch-drop:environment_task_mismatch | 1 | Recorded Python 3.7 cannot exercise the older-Python struct Unicode-format compatibility bug. The selected test checks the helper interface, but an incomplete compatibility fix also passes. Excluded by review, not a NOP false positive. |
| patch-drop:flagged_by_instruction_loop | 7 | Flagged during instruction review and excluded after checking the tests. The instruction asks for a fix to spacing around exec/eval, but the selected test checks different Python 2 formatting behavior. Fixing the reported issue is not enough to pass. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-black-23 Flagged during instruction review and excluded after checking the tests. The test expects one closed event loop to remain in an internal tracking table. A fix that cleans up all closed loops immediately can fail even though it prevents the leak. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-tornado-6 Flagged during instruction review and excluded after checking the tests. The test needs a successful response from the public httpbin.org service. A correct fix can fail when the network or service is unavailable. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-httpie-4 Flagged during instruction review and excluded after checking the tests. The test requires the database query to be written exactly like the reference fix. A different query that finds the same schema and table regardless of letter case can fail. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-luigi-8 Flagged during instruction review and excluded after checking the tests. The test requires the exact command fish --version. A fix that safely gets the shell version another way can fail. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-thefuck-3 Flagged during instruction review and excluded after checking the tests. The tests require environment variables to be read through one particular helper. A correct fix that reads them another way can fail. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-ansible-12 Flagged during instruction review and excluded after checking the tests. The tests require shell aliases to use a particular text layout and order of environment variables. An alias that works the same way but is written differently can fail. Evidence: https://github.com/MLI-lab/OT-next-data/tree/main/data/tasktrove_bugsinpy/flagged_instruction_review_20261008/bugsinpy-original-thefuck-16 |
The published dataset differs from upstream in 489 of the original 500 tasks.
This source uses native records: changed means converted/adapted to Harbor.
Tasks can have multiple change labels, so category counts may overlap.
| Change label | Tasks affected | Explanation |
|---|---|---|
| anti-cheat-instruction | 427 | Added the “Do not cheat…” sentence to the task instructions. |
| upstream-to-harbor | 489 | TaskTrove used GPT-4o-mini to generate simplified buggy code, pytest tests and instructions from upstream fix diffs. A ten-task pilot using GPT-6 Astra as an LLM judge found instruction/test alignment and task-quality problems. We rebuilt the tasks from the original BugsInPy benchmark. For each bug, its metadata names the project, buggy and fixed commits, Python version, test files and test commands. The agent starts with the actual project files from the buggy commit. The patcher turns upstream setup scripts and dependency requirements into working build recipes, adding compatible container images, required build tools and pinned dependencies. The verifier tests come from the upstream fixed commit. The reference solution also comes from that commit and applies the upstream fix to the buggy project. The initial instruction combines a short task template with the title and body of the issue linked from the fixing commit. A separate instruction review and rewrite loop then checks alignment with the verifier, clarity and solution leakage, and rewrites instructions where needed. |
Runtime and validation time
| Setting | Recorded value |
|---|---|
| Cluster | helma |
| Backend | apptainer |
| Architecture | x86_64 |
| Network | host |
| Python | 3.12.14 |
| Harbor | 0.8.1 |
| Apptainer | apptainer version 1.5.4-1.el9 |
| Requested CPUs | 48 |
| Requested memory | 128G |
| Trial concurrency | 8 |
| Static concurrency | 24 |
| terminal-bench commit | 1dcda8716784493721921c23e4bc7f7d988b4494 |
| harbor-validation commit | 7faf878c14b6d72737579ec936ebf5f74ba3194d |
| Unique cached images / tasks kept | 0 / 489 |
| Tasks with cached images | 0 / 489 |
| Stage | Node | Tasks | Tasks at once | Duration | Median per task | 90th percentile |
|---|---|---|---|---|---|---|
| 1 | h31-39 | 489 | 24 | 0:00:07 | 0.3 s | 0.4 s |
Total recorded stage time: 0:00:07. Excludes queue time and work outside these stages.
Static checks excluded by the contract: 13. Recorded but not required: none. Reasons are in the run file.
Image manifest: images/manifests/2026-10-08-bugsinpy-original-fc23c4eacb04.json. Cached images use Apptainer; Dockerfiles remain in the tasks.
Automatically generated data card:
Validation details (bugsinpy-original)
- Native BugsInPy lists 501 bugs. Keras 12 has identical buggy and fixed commits and produces no task payload; this release accounts for the other 500 tasks.
- Build, reference-solution and no-op checks passed on the 489 retained task packages before instruction review. Only instructions changed afterwards; stage 1 was rerun on the final packages. The validation table reports that final stage-1 run.