plawanrath/StableHLO-Spec-30

StableHLO-Spec-30是一个包含30个手工编写的自然语言到StableHLO对的数据集，涵盖10个操作家族。该数据集是NeurIPS 2026论文《Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR》中发布的六个NL→MLIR基准测试之一。数据集格式为每行一个JSON记录，存储在`data/test.jsonl`文件中，包含`dialect`、`difficulty`、`id`、`mlir`、`nl`和`notes`等字段。数据集采用Apache-2.0许可证，主要用于文本生成任务，特别是与MLIR和代码生成相关的任务。

StableHLO-Spec-30 is a dataset of hand-authored natural language to StableHLO pairs across 10 op families (n=30). It is one of six NL→MLIR benchmarks released alongside the NeurIPS 2026 paper *Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR*. The dataset is in JSONL format with one JSON record per line in `data/test.jsonl`, and includes fields such as `dialect`, `difficulty`, `id`, `mlir`, `nl`, and `notes`. Licensed under Apache-2.0, it is intended for text-generation tasks related to MLIR and code-generation, with verifier-valid pass-rate as the primary evaluation metric.