Validation run 2026-10-03-c536bae24db8
CrossCodeEval keeps 6,710 tasks after 1,053 exclusions during the earlier TaskTrove repair. All 6,710 passed this validation run; 0 additional tasks were archived.
| Data source | Before repair | Excluded during repair | Validated and kept | Archived by validation |
|---|---|---|---|---|
| crosscodeeval-csharp-v5 | 1,768 | 415 | 1,353 | 0 |
| crosscodeeval-java-v4 | 2,139 | 244 | 1,895 | 0 |
| crosscodeeval-python-v3 | 500 | 68 | 432 | 0 |
| crosscodeeval-typescript-v3 | 3,356 | 326 | 3,030 | 0 |
Changes
| Change label | Tasks affected | Explanation |
|---|---|---|
| context-and-verifier-repaired | 6,710 | Issue: TaskTrove omitted the benchmark's cross-file context. Its C#, Python and TypeScript verifiers gave 0.25 reward for short code-like answers and could give full reward for prefix or leading-identifier matches without a correct completion. Java used a whole-answer diff that rejected different indentation. Fix: Replace this scoring with normalized exact match of the first statement, add a reference solution, and pin verifier dependencies. For uniquely matched upstream tasks, restore retrieved cross-file excerpts after removing target-file and answer-containing excerpts. Exclude tasks when the supplied code and known library names do not give the agent enough information to infer names required by the expected answer. Also exclude tasks whose scored answer contains no variable, function or other code names. |
The context and verifier repairs were already present in the input to this validation run. See the original TaskTrove repair and review. The input was validated with pinned verifier dependencies.
Earlier exclusions
| Exclusion label | Tasks | Explanation |
|---|---|---|
| Required code names cannot be inferred | 1,043 | The expected answer uses names, such as methods or variables, that the agent cannot infer from the prompt, the available excerpts from other files, or the filter's list of known library names. None of the available sets of excerpts provides enough information. The agent would have to guess these names to match the expected answer. |
| Expected completion contains no code names | 10 | The part of the expected answer that is scored contains no variable, function, class or other code names (identifiers). These tasks were excluded by the patcher's rule requiring at least one such name in the scored completion. |
| Data source | Required code names cannot be inferred | Expected completion contains no code names |
|---|---|---|
| crosscodeeval-csharp-v5 | 415 | 0 |
| crosscodeeval-java-v4 | 242 | 2 |
| crosscodeeval-python-v3 | 68 | 0 |
| crosscodeeval-typescript-v3 | 318 | 8 |
These exclusions preceded this PR. Its published validation record covers the 6,710 retained tasks; it does not record the earlier exclusions as validation failures or contain their original task payloads.
Validation results
| Stage | Passed | Failed |
|---|---|---|
| 1: Static checks | 6,710 | 0 |
| 3: Build, start and setup | 6,710 | 0 |
| 4: Reference solution | 6,710 | 0 |
| 5: No-op | 6,710 | 0 |
Static checks excluded by the contract: 13. These include the separate-verifier requirement (CrossCodeEval grades in the agent container) and test.sh tool-installation checks (the scorer uses the Python standard library). All exclusions and their reasons remain in the linked validation record.
Runtime and validation time
| Setting | Recorded value |
|---|---|
| Backend | apptainer |
| Architecture | x86_64 |
| Network | host |
| Stage | Node | Tasks | Tasks at once | Duration | Median per task | 90th percentile |
|---|---|---|---|---|---|---|
| 1 | h32-32 | 6710 | 16 | 0:25:28 | 0.5 s | 1.4 s |
| 3 | h32-32 | 6710 | 16 | 1:55:38 | 13.4 s | 24.8 s |
| 4 | h32-32 | 6710 | 16 | 1:41:01 | 11.2 s | 27.8 s |
| 5 | h32-32 | 6710 | 16 | 1:47:49 | 11.5 s | 29.8 s |
Total recorded stage time: 5:49:57. Excludes queue time and work outside these stages.