0xanto/crux-data

CRUX decision data

This is the data of the CRUX research project (code: https://github.com/0xAnto/crux). The project tests if rule-conditioned computation and counterfactual learning help a small encoder (about 17M parameters) to decide tasks with rules that it did not see in training. The trained models are at https://huggingface.co/0xanto/crux.

Usage

from datasets import load_dataset

crux = load_dataset("0xanto/crux-data", "crux")  # train, dev_id, validation, test (synthetic)
proofwriter = load_dataset("0xanto/crux-data", "proofwriter", split="test")
boardgameqa = load_dataset("0xanto/crux-data", "boardgameqa", split="test")
dynabench = load_dataset("0xanto/crux-data", "dynabench", split="test")
print(crux["validation"][0]["options"])

Record

Each record is one decision. All configs use the same top-level fields. The meta keys differ between the configs.

Field	Type	Content
id	string	Unique in its file
source	string	The source data set, for example synthetic or proofwriter
task_type	string	binary (True / False / Unknown) or multi_choice
state	string	The facts
instruction	string	The rules, then the question after Question:
options	list of strings	The candidate answers
target	int	The index of the correct option
meta	dict	Source data. Synthetic: the template, the rule composition, the counterfactual pair (pair_id, pair_role, edit_type, label_flip), and the paraphrase group

docs/schema.md and docs/generator_spec.md in the code repository define the record and the generator.

Configs and splits

Config	Split	Records	Content	License
crux	train	75,646	Synthetic counterfactual data from the project generator	Apache-2.0
crux	dev_id	4,162	Synthetic, in-distribution development split (seen rules)	Apache-2.0
crux	validation	10,106	Synthetic, unseen rule compositions. The project uses it for all model selection	Apache-2.0
crux	test	10,086	Synthetic, unseen rule compositions	Apache-2.0
proofwriter	test	32,135	ProofWriter open-world sets: depth5 18,857, birds_electricity 5,270, pararules 8,008 (meta.subset)	CC BY 4.0
boardgameqa	test	15,000	BoardgameQA, 15 subsets of 1,000 (meta.subset)	CC BY
dynabench	test	543	The DynaBench test set of the DynaGuard paper	MIT

The project evaluated the test data one time, in Step 8. Do not use the test data for model selection. Super-NaturalInstructions has no config: data/benchmarks/supernat_ids.jsonl holds the record IDs, the task IDs, and the SHA-256 of each record only (4,100 records, 41 tasks), and supernat_tasks.json holds the license of each task. To get the records, download them from the official source and match them with the hashes.

Integrity

splits_manifest.json gives the SHA-256 of each frozen split file of the project. The frozen test split is one file. Here it is two files, because the two sources have different meta keys. MANIFEST.json lists the two files in order (test_files_in_order). The two files joined in this order give the frozen test split, with the same SHA-256. MANIFEST.json also gives the SHA-256 of every file in this repository.

Attribution and licenses

NOTICE gives the attribution of ProofWriter, RuleTaker, BoardgameQA, DynaBench, and Super-NaturalInstructions. DATA_LICENSES.md records the license check of each source.

Limitations

• The ProofWriter sets are the open-world versions (True / False / Unknown). Compare results only with published open-world results.
• The sentence frames and names of the ProofWriter depth5 set occur in the pretraining data of the Ettin encoders.
• The synthetic data comes from one generator, with fixed templates and vocabularies.
• DynaBench has 543 records, so the confidence intervals on it are wide.

Citation

@misc{ranjith2026crux,
  author       = {Antony Ranjith F.},
  title        = {Does Rule Conditioning Help a 17M Encoder Decide on Unseen Rules? A Negative Result for the {CRUX} Architectures},
  year         = {2026},
  howpublished = {\url{https://github.com/0xAnto/crux}}
}