# crosscodeeval-java-v4

## Original datasource

- CrossCodeEval (Java split) — https://arxiv.org/abs/2310.11248

## Domains

Software Engineering

- **Software Engineering:** Every sampled task asks the agent to complete the next Java statement in real open-source application code (CLI shells, Spring controllers, Android services, JDBC queries, Minecraft mod renderers), using in-file context plus retrieved excerpts from other repository files. The reference is the project's actual next line, e.g. a call to a project-specific API such as Contrast.ratioOfTones or commands.createRequest.

## Agent capabilities

Exploration, Specification Adherence

- **Exploration:** Intended practice: the instruction points to read-only cross-file excerpts under /setup_files/context (4-5 chunked files per task, 1-2.5 KB total) and the correct continuation usually depends on identifiers only visible there (method names, field accessors, constants). The agent must inspect these files before answering. No trajectories were supplied, so observed behavior cannot be confirmed.
- **Specification Adherence:** Intended practice: the output contract requires writing only the missing fragment to /app/solution.txt with no fences, repeated context, or prose, and the verifier (a port of the official cceval metric) awards reward only for exact match of the first statement after comment removal and whitespace normalization. Any extra tokens or formatting drift fails the task. Not verified from trajectories.

## Expected benchmark transfer

These are hypotheses, not measured training gains.

- **SWE-bench Verified random-100**: Both are Harbor-style sandbox tasks run by a terminal agent that must read repository files and produce code consistent with project-specific APIs. Practicing cross-file API inference and precise code edits in an agentic file-writing format should modestly help repository-grounded repair, though the transfer is limited by the language gap (Java vs Python) and the much shorter horizon of single-statement completion.

## 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