plawanrath/MLIR-Functional-Reference-30

MLIR-Functional-Reference-30是一个手写的功能性正确参考集，包含arith、linalg+memref和stablehlo三种方言（每种方言10个实例，共30个实例）。该数据集是NeurIPS 2026评估与数据集轨道论文《Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR》的一部分。数据集格式为每行一个JSON记录，包含多个字段，如canonical_fn_name、dialect、nl等。数据集主要用于验证自然语言到MLIR的转换正确性，是测试专用的，不适合用于微调。数据集遵循Apache-2.0许可证。

MLIR-Functional-Reference-30 is a hand-authored functional-correctness reference set for arith, linalg+memref, and stablehlo (n=30, 10 per dialect). 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*. The dataset format is one JSON record per line, containing fields such as canonical_fn_name, dialect, nl, etc. It is primarily used for verifying the correctness of natural-language→MLIR pairs and is test-only, not suitable for fine-tuning. The dataset is licensed under Apache-2.0.