Datasets:
Scale-SWE-Verified: 16,786 retained tasks and 416 archived
#10
by FWeindel - opened
Keeps 16,786 tasks and archives 416 from PrimeIntellect/Scale-SWE-Verified, covering all 17,202 source tasks.
Includes the full validation run and reruns of repaired tasks.
| Validation | Latest passing outcomes among retained tasks |
|---|---|
| Stage 1: static checks, training profile | 16,786 / 16,786 |
| Stage 3: build and startup | 16,786 / 16,786 |
| Stage 4: reference solution, reward 1 | 16,786 / 16,786 |
| Stage 5: no-op, reward 0 | 16,786 / 16,786 |
The evidence combines the original full audit with targeted repair reruns on Helma using Apptainer. The later full reproduction failed during startup and contributes no passing evidence. This PR does not claim a successful fresh reproduction of the complete final dataset. The training profile’s static exclusions and each historical report’s provenance are retained in the run record and evidence.
Archived tasks
| Reason | Tasks | Explanation |
|---|---|---|
| image-dependency-mismatch | 3 | Installed dependencies conflict with the selected tests. |
| image-missing-test-dependency | 2 | The image lacks a package needed to collect tests. |
| reference-solution-reward-zero | 149 | The reference solution scores zero; task-specific test, setup and verifier causes are recorded in the manifest. |
| task-changes-only-test-files | 15 | Only test files change, leaving no source fix for an agent to make. |
| verifier-collection-error | 55 | The selected tests cannot be collected or set up. |
| verifier-execution-checker-pythonpath | 4 | A test rejects the validation runtime’s PYTHONPATH setting; a runtime limitation. |
| verifier-execution-error | 2 | Verifier execution fails; detailed causes are saved per task. |
| verifier-external-dependency | 2 | Verification depends on an external resource. |
| verifier-network-dependency | 120 | Tests require remote services that are blocked, unavailable or time out. |
| verifier-no-op-passes | 37 | The tests also pass without the reference fix. |
| verifier-privileged-port | 7 | Tests require a privileged port unavailable in the container. |
| verifier-stale-cassette | 18 | Recorded HTTP fixtures no longer match the client requests. |
| verifier-timeout-without-fix | 1 | Verification times out without an accepted repair. |
| verifier-timing-dependent | 1 | The outcome depends on timing. |
Repairs
The patcher changed 1,197 retained tasks. A task can have more than one repair.
| Change | Tasks | Explanation |
|---|---|---|
| instruction-path-clarified | 750 | Clarify paths using observations after task setup and reference execution. |
| verifier-call-time-imports | 5 | Restore required imports at call time. |
| verifier-cpu-affinity | 5 | Size worker pools from allocated CPU affinity. |
| verifier-debug-stream-isolation | 26 | Isolate completion debug streams from pytest’s file descriptors. |
| verifier-exact-test-ids | 5 | Use exact collected test IDs and reject incomplete or skipped required tests. |
| verifier-import-path | 36 | Import the task checkout and its test helpers. |
| verifier-request-cleanup | 1 | Close a test-created request after its assertions. |
| verifier-required-test-skipped | 5 | Remove an unexecutable required test that always skips because the image includes the C++ extension; the remaining required tests cover the change. |
| verifier-test-fixture-restoration | 367 | Restore upstream test-support files omitted by the verifier’s test patch. |
| verifier-worker-cleanup | 6 | Clean up test workers and isolate problematic shutdown behavior. |
Task-specific reasons and original excluded payloads are preserved in the archive and patch manifest.