LogicBench: Logical Chain-of-Thought Training Dataset

Dataset Description

LogicBench is a large-scale dataset for training models on logical and causal reasoning with chain-of-thought (CoT) explanations. The dataset contains 40,000 training examples, each consisting of:

• Facts: A set of 4-10 atomic facts about entities
• Rules: A set of 3-8 Horn-style inference rules (if-then statements with conjunctions)
• Query: A single query statement to prove or refute
• Valid CoT: A step-by-step proof chain that is logically valid
• Near-miss CoT: An intentionally incorrect proof chain that differs by exactly one corruption
• Response: Binary label (1 if query is provable, 0 otherwise)

Dataset Structure

Data Fields

• question (string): Contains the facts, rules, and query formatted as:

  Facts:
  [Entity] is [property] [PERIOD]
  ...
  
  Rules:
  [premise1] [AND] [premise2] [IMPLY] [conclusion] [PERIOD]
  ...
  
  Query:
  [Entity] is [property] [PERIOD]
  

• valid_cot (string): A logically sound chain-of-thought reasoning that correctly determines whether the query follows from the facts and rules.

• near_miss_cot (string): A near-miss reasoning chain with exactly one corruption type:

  • Reverse rule direction
  • Omit one necessary premise
  • Apply rule with mismatched predicate/variable
  • Substitute entity incorrectly
  • Use a rule that is not applicable

• response (int64): Binary label where:

  • 1 = Query is logically entailed (provable)
  • 0 = Query is not entailed (not provable)

Data Splits

• Train: 40,000 examples

Distribution

• Provable queries (response=1): ~65%
• Non-provable queries (response=0): ~35%

Intended Use

This dataset is designed for:

1. Training reasoning models to perform logical inference with step-by-step explanations
2. Evaluating chain-of-thought reasoning capabilities
3. Contrastive learning using valid vs. near-miss reasoning chains
4. Testing robustness to subtle logical errors

Example

{
  "question": "Facts:\nAlice is brave [PERIOD]\nAlice is kind [PERIOD]\n\nRules:\nbrave [AND] kind [IMPLY] helpful [PERIOD]\n\nQuery:\nAlice is helpful [PERIOD]",
  
  "valid_cot": "Step 1: From the facts, we know Alice has the following properties: brave, kind.\nStep 2: Apply rule 'brave AND kind [IMPLY] helpful'.\nStep 3: The premises (brave, kind) are satisfied.\nStep 4: By modus ponens, we derive that Alice is helpful.\nTherefore, QUERY is TRUE.",
  
  "near_miss_cot": "Step 1: From the facts, we know Alice has the following properties: brave, kind.\nStep 2: Apply rule 'brave AND kind [IMPLY] helpful'.\nStep 3: The premises (brave, kind) are partially satisfied (one premise missing).\nStep 4: By modus ponens, we derive that Alice is helpful.\nTherefore, QUERY is TRUE.",
  
  "response": 1
}

Dataset Creation

The dataset was programmatically generated using:

• Forward chaining inference with Horn clauses
• Systematic corruption of valid reasoning chains
• Balanced distribution of provable and non-provable queries
• Validation to ensure consistency between facts, rules, and labels

Citation

If you use this dataset, please cite:

@dataset{logicbench2026,
  title={LogicBench: A Logical Chain-of-Thought Training Dataset},
  author={Amartya77},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Amartya77/LogicBench}}
}

License

MIT License