code2lora/code2lora-data-ood

code2lora-data-ood 是一个用于后截止时间外分布（OOD）评估的数据切片。它包含来自仓库的提交，这些仓库的首次发布严格在Qwen2.5-Coder预训练截止时间之后。该数据集旨在隔离跨仓库的迁移效果，避免训练集记忆的影响，用于评估模型在外部分布数据上的泛化能力。数据布局包括commits.parquet（每提交的元数据和差异/仓库状态嵌入）和qna.parquet（每个断言行，如前缀、目标、行号等），两者通过(repo_id, commit_index, commit_sha)键连接以重建每提交的问答包。模式与其他数据集匹配，并有配套的in-distribution和静态快照变体数据集。code2lora-data-ood is a post-cutoff OOD evaluation split. It consists of commits from repositories whose first release is strictly after the Qwen2.5-Coder pretraining cutoff. Used to isolate cross-repository transfer from training-set memorisation, enabling evaluation of model generalisation on out-of-distribution data. The layout includes commits.parquet (per-commit metadata + diff/repo-state embeddings) and qna.parquet (per-assertion rows such as prefix, target, lineno, ...), both sharing the same (repo_id, commit_index, commit_sha) key for left join to reconstruct per-commit QnA bundles. The schema matches other datasets, with companion in-distribution and static-snapshot variants.