Datasets:
Add CalibForge filtered by stages 3 and 5
Adds Calibforge_filtered_stages_3_5 from AweAI-Team/CalibForge: 5,317 tasks kept and 114 archived, out of 5,431.
| Stage | Result |
|---|---|
| 3: Verifying build | 5,374 kept, 57 excluded |
| 4: Reference solution reward 1 | Skipped, no reference solutions |
| 5: No-op reward 0 | 5,317 kept, 57 further excluded; 13 verifier repairs passed |
Includes the full validation run and reruns of repaired tasks. Includes the three later stage 3 login-script exclusions and omits the later stage 1 fixes.
Stage 3: Verifying build
- 5,079 tasks patched for build failures: corrected Docker-to-Apptainer handling of working directories, variables, temporary inputs, COPY operations, heredocs and multi-stage builds; restored missing directories and task files, including compiled Python files recovered from upstream images.
- 2,969 tasks patched for verifier preparation: installed dependencies in the agent image to reduce repeated downloads. For 2,841 of these tasks, moved installation commands into
tests/setup.shto measure preparation separately from grading. Mean verifier preparation after patching was 4.81 seconds, measured five times per task. A 10-task before-patch pilot took up to 43 seconds before setup failed, so it did not establish a successful before-patch mean. - 57 tasks excluded: 45 database/server ownership failures, 7 temporary-directory state failures, 2 privileged-port failures and 3 tasks whose login scripts repeatedly started background connections and caused commands to time out.
Stage 4: Reference solution reward 1
Skipped because none of the 5,431 tasks provides a reference solution.
Stage 5: No-op reward 0
- 13 verifier failures fixed: in 10 tasks, missing files the agent must create now fail a test instead of crashing test loading. In 1 task, missing pandas is reported the same way. In 2 tasks, an unrepaired PostgreSQL service or Go build now fails tests instead of aborting test setup. All 13 then scored 0 on the no-op run.
- 57 tasks excluded: 4 scored 1 without any agent work, 10 lacked Python for verification, 8 exceeded memory limits, 11 had missing or incompatible libraries, 8 crashed in native code, 7 exceeded the verifier timeout, 4 lacked pytest, 2 had broken shell test scripts, 1 lacked a required starting file and 2 encountered permission errors.
- Total for stage 5: 70 tasks affected, 13 patched and 57 excluded.
Total: 5,317 tasks from 5,431 kept after stages 3 + 5.
Files: README.md, patch.py, tasks.parquet, archive.parquet, manifest.json and run.json. The patcher reproduces the retained task payloads from the pinned source; the archive preserves original excluded tasks and their reasons.
Cached Apptainer images now cover 5,305 of 5,317 tasks (5,198 distinct environments). The remaining 12 tasks require an image build. tasks.parquet contains pinned image references; images.json lists coverage and missing environments.
