license: apache-2.0
language:
- ru
pretty_name: DSTIT
STIT Logic Problems (Xu Axiomatization)
Dataset Description
This dataset contains 1000 synthetically generated problems designed to test logical reasoning capabilities, specifically within the framework of the deliberative 'seeing-to-it-that' (dstit) logic of agency, as axiomatized by Ming Xu (1998) and further analyzed using sequent calculus (G3) in Negri and Pavlović (2020).
Each problem presents a scenario involving agents, their actions (or inactions), and consequences, phrased in natural language (Russian). The core of the problem is based on a specific axiom or logical property discussed in the reference paper. The dataset provides the natural language problem, a hint referring to the relevant logical principle, and the correct answer including a formal logical representation and a reference to the paper.
The dataset is intended for researchers and developers working on logical reasoning in AI, formal methods, natural language understanding of logical statements, and knowledge representation.
Dataset Structure
The dataset is provided in JSON Lines format (.jsonl
). Each line in the file represents a single problem and is a JSON object with the following fields:
description
: A brief description indicating which axiom or logical concept the problem illustrates (e.g., "Применение A1: D^i A ⊃ A").question
: The problem scenario and question posed in Russian natural language.think
: A hint or suggestion guiding the user towards the relevant logical principle or proof idea from the paper.answer
: An object containing the solution:text
: The answer explained in Russian natural language.formal_logic
: The formal logical formula or derivation step from the axiomatization or sequent calculus, using the notation from the paper (e.g.,D^i A ⊃ A
). Propositional variables are represented by letters (A, B, C), and agents by their symbolic IDs (i, k).paper_ref
: A reference to the section, proposition, theorem, or appendix in the Negri & Pavlović (2020) paper where the relevant proof or discussion can be found.
difficulty
: An integer indicating the estimated logical complexity or the level of axioms involved (from 1 to 5, roughly correlating with the order of axioms in the paper).tags
: A list of strings providing keywords for categorization (e.g., "логика агентности", "аксиома_A1", "dstit", "индивидуальная ответственность", "Логика L0").
Dataset Creation
The dataset was created programmatically based on the logical framework presented in:
- Xu, M. (1998). Axioms for deliberative stit. Journal of Philosophical Logic, 27(5), 505–552.
- Negri, S., & Pavlović, E. (2020). Proof-theoretic analysis of the logics of agency: The deliberative STIT. Studia Logica, 28(1), 205–224. https://doi.org/10.1007/s10699-021-09784-3
A Python script was used to generate 1000 examples. The script contains templates for natural language problems and their corresponding formal solutions, tied to specific axioms (A1, A3, A5, A6, A9, A10, A12 from Xu's L0 and L1, as well as the ideas behind AIAn and APCk as discussed in Negri & Pavlović).
The templates are populated using predefined lists of:
- Agents with symbolic IDs (e.g., Иван (i), Робот-X (x)).
- Simple actions (e.g., подписал документ).
- Simple or complex consequences/states.
- Relations between agents (e.g., действовали независимо).
Each of the 1000 problems is generated by randomly selecting one of the supported axiom templates and filling its placeholders with randomly chosen elements from the lists, ensuring logical consistency with the axiom being demonstrated. The difficulty and tags are assigned based on the selected axiom.
Curated Samples
Here are a few examples from the dataset:
{
"description": "Применение A1: D^i A ⊃ A",
"question": "Иван намеренно увидит к тому, что подписал документ. Следует ли, что Иван подписал документ?",
"think": "Аксиома A1 (D^i A ⊃ A): Что агент 'видит к тому, что', то и есть. В G3DSTIT это следует из правила RDi.",
"answer": {
"text": "Да, Иван подписал документ следует. Это прямое следствие аксиомы A1: D^i A ⊃ A. Если агент намеренно добивается A, то A истинно.",
"formal_logic": "D^i A ⊢ A",
"paper_ref": "Раздел 3, доказательство A1"
},
"difficulty": 1,
"tags": [
"логика агентности",
"аксиома_A1",
"dstit",
"индивидуальная ответственность",
"базовый вывод",
"Логика L0"
]
}
{
"description": "Применение A1: D^i A ⊃ A",
"question": "Иван намеренно увидит к тому, что подписал документ. Следует ли, что Иван подписал документ?",
"think": "Аксиома A1 (D^i A ⊃ A): Что агент 'видит к тому, что', то и есть. В G3DSTIT это следует из правила RDi.",
"answer": {
"text": "Да, Иван подписал документ следует. Это прямое следствие аксиомы A1: D^i A ⊃ A. Если агент намеренно добивается A, то A истинно.",
"formal_logic": "D^i A ⊢ A",
"paper_ref": "Раздел 3, доказательство A1"
},
"difficulty": 1,
"tags": [
"логика агентности",
"аксиома_A1",
"dstit",
"индивидуальная ответственность",
"базовый вывод",
"Логика L0"
]
}
{
"description": "Применение A12: D^i ¬D^k B ⊃ i = k (Impossibility of Meta-agency)",
"question": "Иван намеренно добивается того, что Робот-X не намеренно проигнорировал сигнал. Что можно сказать об Иван и Робот-X, предполагая, что они являются *разными* агентами?",
"think": "Это касается 'мета-агентности' - агента, видящего к тому, что делает или не делает другой агент. Смотрите Theorem* 5.2 о невозможности мета-агентности для *различных* агентов.",
"answer": {
"text": "Если Иван и Робот-X являются разными агентами (т.е. i ≠ x), то утверждение 'D^i ¬D^x B' (что Иван намеренно добивается того, что Робот-X не намеренно B) является логически невозможным. Согласно аксиоме A12 и Theorem* 5.2, D^i ¬D^k B влечет i=k. Таким образом, если i ≠ k, формула 'D^i ¬D^k B' является контрлогичной в этой системе.",
"formal_logic": "D^i ¬D^x B ⊃ i = x",
"paper_ref": "Theorem* 5.2, Аксиома A12"
},
"difficulty": 3,
"tags": [
"логика агентности",
"аксиома_A12",
"dstit",
"мета-агентность",
"идентичность агентов",
"ограничения dstit",
"Логика L1"
]
}
{
"description": "Идея AIA2: Независимость агентов",
"question": "Мария и Алексей действовали независимо. Мария намеренно добивается удалить данные, при этом Мария не намеренно добивается включить защиту. Также Алексей намеренно добивается отменить приказ, при этом Алексей не намеренно добивается включить защиту. Могут ли Мария и Алексей *совместно* добиться включить защиту?",
"think": "Подумайте о принципе независимости (AIAn). Если два независимых агента по отдельности не видят к B (т.е. ¬D^i B и ¬D^k B), могут ли они это сделать совместно (т.е. D^{i,k} B)? AIA2 - это частный случай AIAn для n=2.",
"answer": {
"text": "Нет, не могут. Поскольку Мария и Алексей независимы (согласно Diff(m, a)) и каждый из них в отдельности не добивается включить защиту (¬D^m B и ¬D^a B), их совместное действие также не может привести к включить защиту (¬D^{m,a} B). Это следует из аксиомы AIAn.",
"formal_logic": "(¬D^m B & ¬D^a B & Diff(m,a)) ⊃ ¬D^{m,a} B",
"paper_ref": "Раздел 4.3, Аксиома AIAn (особенно AIA2)"
},
"difficulty": 4,
"tags": [
"логика агентности",
"аксиома_AIA2_Idea",
"dstit",
"независимость агентов",
"коллективная ответственность",
"взаимодействие агентов",
"Логика L1"
]
}
@article{Negri2020ProoftheoreticAO,
author={Sara Negri and Edi Pavlovic},
title={Proof-theoretic analysis of the logics of agency: The deliberative STIT},
journal={Foundations of Science},
year={2020},
volume={28},
number={1},
pages={205-224},
doi={10.1007/s10699-021-09784-3}
}
@article{Xu1998AxiomsFD,
author={Ming Xu},
title={Axioms for deliberative stit},
journal={Journal of Philosophical Logic},
year={1998},
volume={27},
number={5},
pages={505-552},
doi={10.1023/A:1004301324615}
}