Datasets:

plawanrath
/

StableHLO-Held-Out-200

license: apache-2.0
language:
  - en
pretty_name: StableHLO-Held-Out-200
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - mlir
  - code-generation
  - compiler
  - constrained-decoding
  - stablehlo
  - held-out
  - programmatic
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.jsonl

StableHLO-Held-Out-200

Programmatic parametric sweep over 7 StableHLO op families × 6 dtypes × 3 shape ranks (n=200, verifier-clean).

Composition

Instances: 200
Format: one JSON record per line in data/test.jsonl
Schema: fields = dialect, difficulty, id, mlir, nl, notes, source
Verifier: stablehlo-opt v1.4.0 (upstream truth) and iree-compile --compile-to=input (substitute, 50/50 concordant on a stratified n=50 sample)
License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions.

Loading

from datasets import load_dataset
ds = load_dataset("plawanrath/StableHLO-Held-Out-200", split="test")
print(ds[0])

Each record is a self-contained natural-language→MLIR pair; verify-valid pass-rate under the dialect's verifier is the primary evaluation metric.

Source format

The JSONL file at data/test.jsonl is the canonical HuggingFace interface. MLCommons Croissant 1.0 metadata (croissant.json) ships alongside the release.

Datasheet

Key points (full Gebru-style datasheet ships with the dataset archive):

All reference MLIR programs are verifier-clean at the time of release.
Hand-authored (no crowdsourcing, no LLM-authored references).
Test-only — fine-tuning on these benchmarks contaminates future evaluation and is explicitly out of scope.

License

Apache-2.0. See LICENSE.