Emulated-Inc/logical-reasoning-training-pool

Logical reasoning training pool

Public logical-reasoning problems with checkable answers, from two datasets whose licences allow commercial use, read at the pinned revisions named below and laid out twice. Every problem is an entailment problem: a block of premises, one conclusion, and whether the premises make the conclusion true, false or neither. Train on either layer or on both.

pool.jsonl

Every source rewritten into one shape, 340951 rows, one JSON object per line.

Field	What it holds
id	a row identifier unique within this file
kind	entailment on every row
premises, conclusion	what is given, and what is asked about
answer	True, False or Uncertain
subject	the label the source gives the row, such as its reasoning depth
source	the name of the source directory the row came from
source_repo, source_revision	the dataset and the revision it was read at
source_subset, source_split, source_id	where the row sits in that set
provenance_class	how the row came to exist
licence	the licence of the source it came from

Rows are deduplicated within and across sources on the normalised problem text, keeping the first row that carries the problem, in the order of the sections below. The file is written one source after another in that same order, so a sequential read of its first rows trains on one source.

The generated source publishes its own label words. They are written into the common vocabulary here: entailment becomes True and not entailment becomes False, and the raw layer keeps the words the source wrote. That source reasons under the closed-world assumption, so its False means the problem's rules do not entail the conclusion rather than that they entail its negation, and it never uses Uncertain. Only the expert-written rows carry all three labels.

sources/

The same data as published, minus only the rows the filtering below removed, 480371 rows, one directory per source, holding the files at the paths, in the format and with the columns its own repository publishes. Nothing else here was renamed, reshaped, relabelled or deduplicated. Use this layer if you want a field the rewritten one drops, such as the first-order-logic formalisations, or if you would rather map the labels yourself.

The sources

sources/folio

Expert-written first-order-logic entailment problems, each a block of premises in plain English, one conclusion, and a label saying whether the premises make the conclusion true, false or neither, with a parallel formalisation of both in first-order logic. From tasksource/folio at revision 295b95fb4fe9be4ff3f933b73142d142cf6b2c97, files folio_v2_train.jsonl. 219 rows here, of which 219 also appear in pool.jsonl. Provenance class human, licence cc-by-sa-4.0. Its own fields are story_id, premises, premises-FOL, conclusion, conclusion-FOL, label and example_id.

Licence. The licence is read from the LICENSE file of the authors' project repository, https://github.com/Yale-LILY/FOLIO, which is the Creative Commons Attribution-ShareAlike 4.0 International text (the repository API reports CC-BY-SA-4.0, read 2026-09-23).

Worth knowing. The project repository licenses the work under Creative Commons Attribution-ShareAlike 4.0 while the authors' own dataset page tags it mit, and the stricter of the two is what the pool composes against. Several rows share one block of premises and differ only in the conclusion, which is how the set is built, so a random split of these rows is not an independent split.

sources/ruletaker

Generated theories of facts and rules in plain English with a statement to judge against each, under the closed-world assumption, so a statement is either entailed by the theory or not, at reasoning depths from zero to five plus a hand-written natural-language variant. From tasksource/ruletaker at revision a3e0880baeb6ec3d478f4c4d85afe04b21b6cf7f, files data/train-00000-of-00001-52adaa842dd7ed92.parquet. 480152 rows here, of which 340732 also appear in pool.jsonl. Provenance class rule-generated, licence apache-2.0. Its own fields are context, question, label of entailment or not entailment, and config.

Licence. The licence is read from the authors' project repository, https://github.com/allenai/ruletaker, to which the official dataset page https://allenai.org/data/ruletaker redirects. Its LICENSE file is the Apache License 2.0, and its README links the dataset download and closes with the notice Copyright 2020 Allen Institute for AI, Licensed under the Apache License, Version 2.0 (read 2026-09-23). The release archive rule-reasoning-dataset-V2020.2.5.zip itself ships no licence file. The copy read here tags itself apache-2.0 and was checked against that archive: 435199 of its 480152 train rows match an official train row's question and its theory, sentence for sentence, in any order, none matches the later ProofWriter release alone, and every sentence of a 2000-row sample of the remaining natural-language rows appears in the official archive.

Worth knowing. Closed world, so its not entailment label covers both a false statement and one the theory says nothing about, and the rewritten layer writes it as False. The rewritten layer therefore uses only two of the three labels for these rows, and a model trained on them alone never sees Uncertain. Many rows share one theory and differ only in the question. The configurations also repeat one another: the depth-3ext and depth-3ext-NatLang rows largely restate depth-3 problems and the NatLang rows restate depth-3ext-NatLang ones, so about three raw rows in ten are duplicates that the rewritten layer drops.

Provenance and licences

One source is problems written by domain experts. The other is the output of a program that builds a problem and its answer together, which is the provenance class rule-generated and which is most of the rows. Nothing in the pool was generated by a language model, and no answer anywhere in it was produced by a model: every key is either a human annotation or the generator's own record of the answer it built the problem around.

The pool as a whole is offered under cc-by-sa-4.0, which is the most restrictive term its sources compose to. The sources themselves are cc-by-sa-4.0 for the expert-written set and apache-2.0 for the generated set, and both allow commercial use. Each rewritten row carries its own licence in the licence field and each directory under sources/ is one source, so a subset under a single licence can be selected. Attribution goes to the authors named in each source's own citation, and the share-alike term of the expert-written set applies to anything built from its rows.

Filtering

Rows whose text duplicated or overlapped the text of a problem in a held-out collection were removed before publication, from both layers alike, by an exact whole-text match (500 rows) followed by a word 8-gram overlap check (282 rows). Both checks read the premises and the conclusion. An 8-gram carried by more than one held-out problem is a stock phrase rather than a problem, so it was left out of that check, and the 86 rows whose only overlap was such a phrase were counted and kept. The held-out collection is not distributed here. Nothing else was filtered: no reasoning depth, no subject and no difficulty was selected for or against, so the pool still holds the easiest depth-zero generated rows, and the subject field is how to select what you want. That field is the source's own configuration label: depth-0, depth-1, depth-2, depth-3, depth-3ext, depth-5, NatLang and depth-3ext-NatLang for the generated rows.

Two warnings a reader should carry into training. Most rows come from one generator, so their wording repeats heavily and a model can learn the surface pattern rather than the reasoning. And both sources build many rows on one block of premises, differing only in what is asked, so a random split of these rows does not give independent train and held-out halves.