crosscodeeval_v1/README.md

CrossCodeEval

Dataset stats

• Original tasks: 7,763
• Kept: 6,710
• Changed: 6,710
• Archived: 1,053
• Images: 3 original definitions; 3 now, as cached Apptainer bundles.

Dataset characteristics

Complete the next statement in a real C#, Java, Python or TypeScript repository using excerpts from related files.

Domains

Software Engineering: cross-file code completion, repository APIs and library usage.

Capabilities

Exploration: find required names and calling conventions in supplied context. Specification Adherence: write the requested continuation in the required output file.

Intended benchmark transfer: Python subset: HumanEval+, MBPP+ and DS-1000 mini-200. Java/Python/TypeScript: repository exploration for SWE-bench Verified random-100. Labels follow CLI-Universe.

Per-stage breakdown

Patcher rules archived 1,053 tasks before validation: 1,043 with required identifiers absent from the available context and 10 whose scored answer contained no identifiers.

• Build and setup: 6,710 changed, 0 archived. Pinned verifier dependencies; Docker definitions converted to Apptainer for the Marin Harbor bridge and repository runtime patches.
• No-op baseline (reward 0): 6,710 changed, 0 archived. Scoring repaired to reject incomplete prefix matches and normalize comments/whitespace.
• Reference solution (reward 1): 6,710 changed, 0 archived. Reference solutions added.
• Rule-based checks: 6,710 changed, 0 archived. Context guidance added and cross-file excerpts restored with answer leakage removed. No kept task has advisory warnings for relative paths. Every instruction also ends with "You have N seconds to complete this task. Do not cheat by using online solutions or hints specific to this task." (N is the agent timeout); this sentence is not counted as a change.

Results are from the historical validation; kept task payloads are unchanged.

Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution).

Runtime and validation time

• Versions: Apptainer 1.5.4, Python 3.12.14, Harbor 0.8.1.
• Implementation: Marin Harbor 7faf878c14b6 + our runtime patches.
• Resources: Helma; 48 CPU cores, 128 GiB RAM per job, 16 concurrent tasks, no GPUs.

Check	Mean per task	Median per task	Tasks measured
Rule-based checks	3.36 s	0.54 s	6,710
Build and setup	16.50 s	13.36 s	6,710
Reference solution	14.39 s	11.20 s	6,710
No-op baseline	15.35 s	11.50 s	6,710

Means and medians cover check execution, environment start/inspection/stop, and reference/no-op trials respectively. Image builds are excluded.