LogicSkills — Training Data

Companion training releases for the LogicSkills benchmark (EMNLP 2026 Findings). This Hugging Face dataset has three subsets: symbolization, countermodel, and validity, each with a single train split. Each subdirectory contains the corresponding data and task documentation. Code for our data generation pipeline can be found here.

Dataset	Path	Size	Task
Symbolization	symbolization/	100k × 2 languages	sentence → FOL
Countermodel	countermodel/	100k	invalid FOL → finite countermodel
Validity	validity/	100k × 2 languages	which conclusions follow (if any)

Loading the dataset

from datasets import load_dataset

repo_id = "mainlp/LogicSkills"
symbolization = load_dataset(repo_id, "symbolization", split="train")
countermodel = load_dataset(repo_id, "countermodel", split="train")
validity = load_dataset(repo_id, "validity", split="train")

The symbolization and validity subsets each combine English and Carroll data in their train split (200,000 rows per subset). Use the language field to select a language, for example:

english_symbolization = symbolization.filter(lambda row: row["language"] == "english")

The countermodel subset contains 100,000 rows.

Relationship to the benchmark

• These files are for training / fine-tuning / analysis.
• The paper’s evaluation sets live separately (see the paper and the LogicSkills benchmark release). Overlapping benchmark situations are held out of these training dumps.
• Do not treat this release as a drop-in replacement for the benchmark splits.

Languages

• English — controlled English over a fixed lexicon
• Carroll — Carroll-style nonce wording with the same logical forms (language / art-like pairing via pair_id where present)

License & citation

CC-BY-4.0. If you use this data, please cite:

@inproceedings{rabern2026logicskills,
  title     = {LogicSkills: A Structured Benchmark for Formal Reasoning in Large Language Models},
  author    = {Rabern, Brian and Mondorf, Philipp and Plank, Barbara},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026},
  publisher = {Association for Computational Linguistics},
  url       = {https://arxiv.org/abs/2602.06533},
}