plawanrath/StableHLO-Held-Out-200

StableHLO-Held-Out-200是一个程序化的参数扫描数据集，覆盖了7个StableHLO操作家族、6种数据类型和3种形状等级（共200个实例，经过验证器清理）。该数据集是NeurIPS 2026评估与数据集轨道论文《无需重新训练的跨方言泛化：MLIR的模式派生约束解码基准与评估》（匿名提交）发布的六个NL→MLIR基准之一。数据集包含200个实例，格式为每行一个JSON记录，存储在`data/test.jsonl`中。每个记录都是一个自包含的自然语言→MLIR对，验证器有效通过率是主要评估指标。数据集采用Apache-2.0许可证，无第三方知识产权限制。

StableHLO-Held-Out-200 is a programmatic parametric sweep over 7 StableHLO op families, 6 dtypes, and 3 shape ranks (n=200, verifier-clean). 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 200 instances, formatted as one JSON record per line in `data/test.jsonl`. Each record is a self-contained natural-language→MLIR pair, with verify-valid pass-rate under the dialects verifier as the primary evaluation metric. The dataset is licensed under Apache-2.0 with no third-party IP restrictions.