plawanrath/Linalg-Spec-30

Linalg-Spec-30是一个包含30个手工编写的自然语言到MLIR（多级中间表示）对的数据集，用于评估跨方言泛化能力。该数据集是NeurIPS 2026评估和数据集轨道论文的一部分，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。所有MLIR程序在发布时都经过验证器验证，且为单作者手工编写，不包含众包或LLM生成的内容。数据集主要用于测试，不适合用于微调以避免未来评估的污染。

Hand-authored NL→MLIR pairs for linalg named ops under memref semantics (n=30). This dataset is one of six NL→MLIR benchmarks released alongside the NeurIPS 2026 Evaluations & Datasets track paper *Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR* (anonymous submission). The dataset contains 30 instances, each in JSON format with fields such as dialect, difficulty, id, mlir, nl, and notes. All reference MLIR programs are verifier-clean at the time of release, hand-authored by a single author, and intended for test-only use to avoid contamination of future evaluations.