plawanrath/MLIR-Spec-150

MLIR-Spec-150数据集包含150个手工编写的自然语言到MLIR（多级中间表示）对的集合，专注于arith、func和memref三种MLIR方言。该数据集是NeurIPS 2026评估与数据集轨道论文的一部分，旨在评估跨方言泛化能力。数据集格式为每行一个JSON记录，包含方言、难度、ID、MLIR代码、自然语言描述和注释等字段。数据集主要用于验证MLIR程序的通过率，且所有参考MLIR程序在发布时都经过验证器验证。数据集是测试专用的，不适合用于微调。

MLIR-Spec-150 is a dataset of 150 hand-authored natural language to MLIR (Multi-Level Intermediate Representation) pairs for the arith, func, and memref dialects. It is part of a suite of benchmarks released alongside a NeurIPS 2026 Datasets and Benchmarks track paper on cross-dialect generalization in MLIR. The dataset is formatted as one JSON record per line, with fields including dialect, difficulty, ID, MLIR code, natural language description, and notes. It is designed for verifying the pass-rate of MLIR programs under the dialects verifier, with all reference MLIR programs being verifier-clean at release. The dataset is intended for testing only and is not suitable for fine-tuning.