CRUSTPairs: A Cross-Language Dataset of Paired C and Rust Solutions

1857 pairs of C and Rust implementations of the same competitive programming problem, drawn from IBM Project CodeNet (839 pairs, AtCoder) and xCodeEval (1018 pairs, Codeforces). Both sides of a pair are accepted submissions to the same problem, so functional equivalence is inherited from provenance rather than verified by execution. Each record carries the upstream problem statement with its format declared, both implementations with entry points normalised to <code>solution</code>, the source corpus, an SPDX licence expression, a difficulty tier with the raw cross-language embedding similarity it was cut from, and a frozen train/validation/test assignment stratified by source and tier. Licensing is per record, not per dataset. The aggregate is non-commercial because the xCodeEval-derived pairs are CC BY-NC 4.0; the CodeNet-derived pairs are CDLA-Permissive-2.0 and can be filtered out as a commercially usable 839-pair subset. The accompanying code is MIT. See DATA_LICENSES.md in the archive. The deposit also contains the construction and evaluation scripts, the retrieval results reported in the paper, and the LoRA adapter trained on the dataset.