code2lora/code2lora-data-snapshots

该数据集名为code2lora-data-snapshots — static-snapshot RepoPeftBench，专门用于直接投影Code2LoRA训练器。数据集中，每一行对应一个（仓库，锚点提交）的组合。每个提交行包含一个2048维的Qwen3-Embedding仓库状态向量；每个QnA行则是从该快照中保留的测试文件中提取的单个断言。数据集布局包括commits和qna目录，其中commits目录包含不同分割（如train、ir_val、ir_test、cr_val、cr_test、ood_test）的锚点提交快照，qna目录则包含与每个提交行匹配的断言行。分割直接按论文套件命名，无需额外过滤。

The dataset is named code2lora-data-snapshots — static-snapshot RepoPeftBench and is designed for the direct-projection Code2LoRA trainer. Each row corresponds to a (repository, anchor commit) pair. Each commit row includes a 2048-dimensional Qwen3-Embedding repository state vector; each QnA row is a single assertion extracted from a held-out test file at that snapshot. The layout consists of commits and qna directories, with the commits directory containing anchor commit snapshots for different splits (e.g., train, ir_val, ir_test, cr_val, cr_test, ood_test), and the qna directory containing matching assertion rows for each commit row. Splits are named directly by paper suite, requiring no extra filtering.