exp_rle_minimal_instructions-v3
收藏资源简介:
exp_rle_minimal_instructions-v3是一个基于laion/exp_rle_minimal_instructions-v2数据集进行修补的版本,旨在通过改进过滤逻辑解决v2中发现的故障问题。该数据集主要用于代码生成和自动化测试任务,包含结构化任务数据,每个样本以gzipped tar压缩包形式存储,内含instruction.md(任务说明)、task.toml(配置)、environment/Dockerfile(环境定义)、tests/*(测试文件)和solution/*(解决方案文件)。数据集中添加了三个新的过滤步骤:Fixture pass(通过AST解析检查pytest fixture的可用性)、Deep-submodule blocklist(阻止导入已知移除的私有子模块,如pandas、numpy等包的特定子路径)和Niche-package blocklist(阻止导入容器pip安装集之外的niche包,如pipelinewise、jasmine等)。数据规模方面,v2版本包含699行数据,v3的具体行数可通过tasks.parquet元数据查看。该数据集适用于代码生成模型训练、自动化测试评估和任务解决基准测试等场景。
exp_rle_minimal_instructions-v3 is a patched version of the laion/exp_rle_minimal_instructions-v2 dataset, designed to address issues found in v2 by improving filtering logic. This dataset is primarily used for code generation and automated testing tasks, containing structured task data. Each sample is stored as a gzipped tar archive, including instruction.md (task description), task.toml (configuration), environment/Dockerfile (environment definition), tests/* (test files), and solution/* (solution files). Three new filtering steps have been added to the dataset: Fixture pass (checking pytest fixture availability via AST parsing), Deep-submodule blocklist (blocking imports of known removed private submodules, such as specific subpaths of packages like pandas and numpy), and Niche-package blocklist (blocking imports of niche packages outside the container pip installation set, such as pipelinewise and jasmine). In terms of data scale, v2 contains 699 rows, and the specific row count for v3 can be viewed through the tasks.parquet metadata. The dataset is suitable for scenarios like code generation model training, automated testing evaluation, and task-solving benchmark testing.




