crosscodeeval-typescript-v3
Original datasource
- CrossCodeEval (TypeScript split) — https://arxiv.org/abs/2310.11248
Domains
Software Engineering
- Software Engineering: Every sampled task is repository-level code completion in real TypeScript application code (NestJS services, Kysely-style query builders, game screens, browser-extension pages, subtitle parsers). The agent must emit the 1-4 line continuation of a file cut off mid-statement so that it matches the project's existing APIs and types; no debugging, data work, or system administration is involved.
Agent capabilities
Exploration, Specification Adherence
- Exploration: Intended practice: the target identifiers (e.g. setScreen, instantiateNoReturns, BlockTuneConstructorParameters, savePreferences) are not in the truncated prompt but appear in the 4-5 retrieved cross-file excerpts under /setup_files/context/, so the agent must read that directory before committing an answer. No trajectories were supplied, so observed behavior cannot be confirmed.
- Specification Adherence: Intended practice: the verifier (a port of amazon-science/cceval eval_metric.py) gives reward 1 only on exact line-level match of the first statement after comment stripping, and the instructions insist on writing only the raw continuation to /app/solution.txt with no re-emitted snippet, imports, fences, or labels, via the quote-safe write_solution helper. Any drift in output format or extra text yields zero reward. Not observed in trajectories.
Expected benchmark transfer
These are hypotheses, not measured training gains.
- SWE-bench Verified random-100: The practiced skill is producing code that is consistent with APIs, types, and conventions defined in other files of the same repository, discovered by reading retrieved cross-file context inside a Terminus-2 container. That repository-grounded completion habit is the closest match in the suite, although transfer is expected to be modest because the dataset is TypeScript-only and single-statement while SWE-bench is Python issue repair with executable tests.
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