hybrid-diff-ar/stack-v2-sparse-classes-36k

这是一个包含36,000个样本的快照数据集，用于Diffusion + Autoregressive混合代码生成实验。数据来源于`bigcode/the-stack-v2-dedup`的Python子集，通过AST级别的类过滤提取。数据集分为训练集（35,000个样本）、验证集（500个样本）和测试集（500个样本）。每个样本是一个Python类，包含自然语言提示、类/方法签名和文档字符串、方法体、完整文本、重构的类代码以及源元数据等信息。过滤条件包括每个类有2到6个方法，每个方法都有非空文档字符串，方法体有3到30行非空代码，重构的类能够解析为Python AST，并排除了测试/文档/示例/供应商/生成的文件。

This is a 36,000-sample snapshot for Diffusion + Autoregressive hybrid code generation experiments. The data is extracted from `bigcode/the-stack-v2-dedup`, Python subset. The extraction uses Stack v2 metadata as source of truth, groups candidates by `repo_name + revision_id`, fetches files with git partial fetch + sparse checkout, then applies AST-level class filters. The dataset is split into training (35,000 samples), validation (500 samples), and test (500 samples) sets. Each sample is a Python class with fields such as natural-language prompt, class/method signatures and docstrings, method bodies, full text, reconstructed class code, and source metadata. Filters include 2 to 6 methods per class, non-empty docstrings for every method, 3 to 30 non-empty lines per method body, reconstructed classes that parse as Python AST, and exclusion of tests/docs/examples/vendor/generated files.