drotherm/whetstone-c18

C18: PrOntoQA deductive entailment

First-pass snapshot v1: generate once, validate, clean, and retain the survivors. The generators have not been redesigned to increase structural diversity. Each configuration has one pool split. Consumers own experimental partitions. Keep rows with the same group_id together when partitioning. Private evaluation gold is deliberately included in this public dataset. Render target prompts from prompt_inputs only.

Configuration	Candidates	Retained	Groups
default	10000	10000	10000
hard	3750	3750	3750

from datasets import load_dataset

data = load_dataset(
    "drotherm/whetstone-c18", "default", split="pool", revision="v1"
)

For an experiment, resolve v1 and pin the full immutable Hub commit SHA. A future snapshot v2 may change the generators and retained populations.

Fields

• id: native instance identifier; unique within each configuration.
• seed: generation seed. Several rows can share it; it is not a grouping key.
• categories: ordered task category labels.
• prompt_inputs: a struct of named public string inputs, preserved verbatim.
• gold: private frozen answer or, for C22, serialized constraint stack.
• group_id: v1 SHA-256 group identity under the documented cleaning policy.

Scoring and cleaning

Extract the final True/False verdict using C18's scorer, then exact match.

default

Duplicate policy: sentence-order-and-whitespace. Removed counts: {}. Category counts: {"D1": 2500, "D2": 2500, "D3": 2500, "D5": 2500}. Limitations: Alpha-equivalent reasoning graphs remain. Category balance is not restored after cleaning. See default/audit.json for file hashes and default/rejections.jsonl for removed IDs/reasons.

hard

Duplicate policy: sentence-order-and-whitespace. Removed counts: {}. Category counts: {"D10": 1250, "D5": 1250, "D8": 1250}. Limitations: Alpha-equivalent reasoning graphs remain. Category balance is not restored after cleaning. See hard/audit.json for file hashes and hard/rejections.jsonl for removed IDs/reasons.

First-pass generation was capped at 5 completed batches for turnaround time. Reproduce with --preset hard --max-batches 5.

Configurations share 0 cleaning groups. Keep matching group IDs together across configurations; hard is a difficulty preset, not an independent held-out split.

Provenance and reproduction

Generators and graders: whetstone-envs at 1a7dcf3f7f6eaee133f8edbfa2748c884ee51fbd. The sampling and cleaning scripts are included under reproduction/. Copy them into scripts/datasets/ in that source checkout. Run with Python 3.12, the locked C18 extra, and pyarrow==26.0.0 / huggingface-hub==2.2.0.

uv run --extra c18 --with pyarrow==26.0.0 --with huggingface-hub==2.2.0 \
  python scripts/datasets/generate.py --task c18 --output /path/to/raw
uv run --extra c18 --with pyarrow==26.0.0 \
  python scripts/datasets/clean.py \
  --source /path/to/raw/c18/default --output /path/to/clean/default

Use --preset hard for the hard configuration. C18 uses bounded 250-example batches per depth with recorded fresh seeds. Each configuration targets approximately 10,000 raw candidates.

Category balance is not restored after cleaning. Counts are not a claim of independent samples. Attribution and applicable licenses accompany this dataset.