ZijiantheJack's picture
Add DeepSWE coding trajectory, reference review and patch triplets
20b3c2a verified
|
Raw History Blame Contribute Delete
4.88 kB
# Core artifact provenance and scope
[Artifact home](../README.md)
## Source revisions
The core payload was curated from artifact revision
`daf21c455b0008e49a9a706fabdb4951a449352c`. That revision traced back to the
framework and analytical-source revisions recorded in `artifact/origin.json`.
## What was selected
Selection combines executed numerical dependencies with explicit coverage of:
- 100 Stage 1 patch pairs, tasks, human/generated references, and evaluations;
- 339 Stage 2 implementation/reference patch pairs, coding trajectories,
generated candidate/reference comments, and official verification records;
- 4,068 selected review outputs, native judgments, prompts, usage, and timing;
- 678 additional full-trajectory reviews, their selected-attempt evidence and
native judgments, and separately preserved API-aligned Diff-only judgments;
- only the native evaluation, reference-generation, analysis code, and tests.
DeepSWE task definitions, hidden test patches, full test-source attribution
packets, task Dockerfiles, benchmark repositories, image layers, manuscript
files, unrelated experiments, and bulk reviewer logs are excluded.
## Data representation and execution
Scientific input content is preserved. Identical files use internal relative
aliases, with both alias target and content checked. `artifact/research-files.json`
records source hashes, deduplication, and
any explicit projections or adaptations.
Directories are organized by stage, data role, coding model, reviewer and
method. `research/source-locations.json` binds locations appearing inside
captured records to current files. Raw trajectories, prompts, comments,
judgments and verifier outputs keep their recorded contents. Runtime entrypoints
and derived source bindings use the descriptive layout.
| Component | Representation or behavior |
| :--- | :--- |
| Codex session exports | Original metadata headers identify sessions; the 42 duplicate-usage proof sessions include complete logs. |
| Stage 1 metadata | PR identifiers, title, and body provide the context consumed by native evaluation. |
| Map inventory | Frozen log-presence records, source-bound case logs, and method representations support map analysis. |
| Severity replay | Frozen accepted labels and native coverage determine the reported strata; benchmark test-source packets are external inputs for fresh labeling. |
| Stage 2 generation | An explicit `--config` selects separate fresh-run inputs and outputs. |
| Native dependencies | The optional SDK dependencies support model-based evaluation and reference generation. |
Result replay uses the native scorer, matching/coverage rules, selected
attempts, comments, official outcomes, and recorded prices.
`artifact/reproduction-baseline.json` records the observed replay and export
hashes.
Captured source manifests and protocol hashes can mention omitted files.
They describe the recorded experiment. The core indices and current inventory
describe the distributed files; no empty stand-ins for task bodies or omitted
reviewer logs are created.
The full-trajectory records come from the completed
`diff-full-trajectory-20261001` experiment. Captured study, API-alignment plan,
source hashes, final state, paired task records, and API-aligned results are
preserved unchanged under `research/stage2/analysis/full-trajectory/`.
The main `results.json`/`results.csv` in that directory are recomputed with the
original frozen Diff-only baseline used by the paper. The archived API-aligned
comparison remains separately named; no captured scores are rewritten.
Exact full-trajectory input hashes are verified after reconstruction from the
already distributed complete coding trajectories.
New capture-machine paths are mapped in the supplement's own
`source-locations.json`, preserving the original six-method catalog and its
frozen analytical provenance.
## Integrity
`artifact/manifest.json` seals the complete non-derived payload, including
relative alias targets and content hashes. `results/`, dependencies, and
externally supplied repositories are not published inputs. All data aliases
resolve to files inside `research/`; source-location mappings do not require
compatibility directories.
```sh
python3 -I -S tools/verify-artifact.py --secrets
```
Credential-pattern checks emit only paths/categories, never matching values.
No authentication stores or live usage database are included.
## Rights
The native framework package declares ISC. That declaration does not
relicense benchmark data or upstream project code.
DeepSWE's [Apache-2.0 notice](../LICENSES/DeepSWE-Apache-2.0.txt) and
[upstream attribution table](../LICENSES/UPSTREAM-PROVENANCE.md) are retained
without the task repository. The task patches remain subject to their
upstream projects' terms. No blanket redistribution clearance, DOI, or
artifact-evaluation badge is implied.