iac-eval-v2
收藏资源简介:
IaC-Eval v2 是一个现代化的 Terraform 代码生成基准测试数据集,是 NeurIPS 2024 IaC-Eval 基准的更新和扩展版本。该数据集专门用于评估大型语言模型在生成符合特定策略约束的云基础设施即代码(IaC)程序方面的能力。数据集包含 186 个经过验证的任务(目前仅限 AWS 环境),每个任务都配有自然语言描述、结构化约束意图、用于评分的 OPA Rego v1 策略、黄金标准的 HCL 参考输出以及所需的 Terraform 和 OPA 版本信息。任务难度分为 1(简单)到 6(专家)六个等级,覆盖了 34 种不同的 AWS 服务。评分机制是确定性的,依赖于 `terraform plan` 和 `opa eval` 的通过与否,而非使用 LLM 作为评判者。数据集采用扁平 CSV 格式,旨在为代码生成模型,特别是在云基础设施自动化领域,提供一个可靠、可复现的评估基准。
IaC-Eval v2 is a modern Terraform code generation benchmark dataset, serving as an updated and expanded version of the NeurIPS 2024 IaC-Eval benchmark. This dataset is specifically designed to evaluate the capability of large language models (LLMs) in generating cloud Infrastructure as Code (IaC) programs that comply with specific policy constraints. The dataset contains 186 validated tasks (currently limited to AWS environments), where each task is accompanied by natural language descriptions, structured constraint intentions, OPA Rego v1 policies for scoring, gold-standard HCL reference outputs, and required Terraform and OPA version information. The tasks are divided into six difficulty levels ranging from 1 (simple) to 6 (expert), covering 34 distinct AWS services. The scoring mechanism is deterministic, relying on the pass or fail results of `terraform plan` and `opa eval` rather than employing LLMs as evaluators. The dataset adopts a flat CSV format, aiming to provide a reliable and reproducible evaluation benchmark for code generation models, particularly in the domain of cloud infrastructure automation.





