crosscodeeval-python-v3
Original datasource
- CrossCodeEval — https://arxiv.org/abs/2310.11248
Domains
Software Engineering
- Software Engineering: Every sampled task is Python code completion inside a real repository file: the prompt ends mid-statement (e.g.
assert client.,grad_norm = utils.,subtask_ids: List[str] = config.) and the agent must write the next statement, typically an API call or attribute access whose name and signature are defined in other repository files supplied as retrieved cross-file context under /setup_files/context/. The verifier is a port of the cceval metric code and rewards exact match of the first completed statement after comment removal and whitespace normalization.
Agent capabilities
Exploration, Specification Adherence
- Exploration: Intended practice: the correct continuation depends on identifiers that are not visible in the prompt (e.g.
add_rand_var(...),get_grad_norm_,smc_steer,bet_amount/raise_limit), so the agent is expected to open the 4-5 retrieved context excerpts in /setup_files/context/ to discover the right API names and argument conventions before answering. No trajectories were supplied, so this is inferred from the task design rather than observed behavior. - Specification Adherence: Intended practice: the instruction imposes a strict output contract (write only the short continuation to /app/solution.txt via
write_solution, no re-emitted snippet, no imports, fences, or labels) and the verifier scores exact match of the first statement, so any formatting drift or extra text yields reward 0. Edit similarity and identifier F1 are logged but do not affect the pass/fail reward. Inferred from task design; no trajectories were supplied.
Expected benchmark transfer
These are hypotheses, not measured training gains.
- HumanEval+: Both require producing syntactically and semantically correct Python that completes a partial program; practice at emitting exact, idiomatic Python continuations with correct argument lists and attribute chains should modestly help zero-shot Python function completion.
- MBPP+: Same Python code-writing skill as above: short, precise, correct Python statements using standard library and common third-party APIs (os.path, unittest assertions, torch tensors) that appear in the sampled tasks.
- DS-1000 mini-200: DS-1000 problems are largely about completing a partial Python snippet with the right library call, and the campaign runs them through Terminus-2 in Harbor, the same agentic file-writing setup used here. The sampled tasks include numpy/torch/data-loader API completions (e.g.
loss.item(),sim_matrix), so both the skill and the delivery format overlap. - SWE-bench Verified random-100: The dataset trains reading related repository files to infer how a project's own classes and helpers are used (e.g. hotpot, genbench, pyvsc, appsignal internals) before editing code, which is a core sub-skill of repository-level issue repair in a terminal agent. Expected gain is modest since the tasks here are single statements with no test execution.
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