crosscodeeval-csharp-v5
Original datasource
- CrossCodeEval (C# split) — https://arxiv.org/abs/2310.11248
Domains
Software Engineering
- Software Engineering: Every sampled task is cross-file code completion in a real C# repository (Unity mods, ASP.NET extensions, WPF samples): the agent must produce the next statement of a cut-off declaration, consistent with types, attributes, and method signatures defined in sibling files supplied under /setup_files/context.
Agent capabilities
Exploration, Specification Adherence
- Exploration: Intended practice: the correct continuation (e.g. 'LogLevel> LootingLogLevels;', 'StyleSheetName => nameof(SceneSelectorWindow);', 'LogdashboardAccountAuthorizeFilter filter, ...') depends on identifiers declared only in the retrieved cross-file excerpts, so the agent is expected to read those files before answering. No trajectories were supplied, so observed behavior cannot be confirmed.
- Specification Adherence: Intended practice: the verifier (a port of the official cceval metric) truncates the prediction to its first statement and requires an exact line-by-line match with the reference. The instruction further requires writing only the short fragment to /app/solution.txt via a quote-safe helper, with no re-emitted prefix, usings, code fences, or labels. Output drift of any kind fails the task. Not observed in trajectories since none were supplied.
Expected benchmark transfer
These are hypotheses, not measured training gains.
The tasks are single-statement C# completions scored by exact string match against a fixed reference, with no compilation or execution. The suite's code benchmarks are Python-centric and execution-verified (HumanEval+, MBPP+, DS-1000, CRUXEval, SWE-bench Verified), and the terminal benchmarks (Terminal-Bench 2.0, OpenThoughts-TBLite) require multi-step shell work, whereas these tasks reduce to reading a few context files and writing one line. The language, output format, and verification signal are too far from any suite benchmark to reasonably expect a score increase.
Files
tasks.parquet: kept tasks.archive.parquet: excluded tasks with validation reasons.
Generated with claude-fable-5-1; domain and capability labels follow CLI-Universe. Taxonomy source: https://arxiv.org/abs/2606.22883v1