metadata

license: cc-by-sa-4.0
dataset_info:
  features:
    - name: instruction
      dtype: string
    - name: context
      dtype: string
    - name: question
      dtype: string
    - name: targets
      sequence: string
    - name: target_choices
      sequence: string
    - name: target_scores
      sequence: int32
    - name: reasoning
      dtype: string
    - name: source_data
      dtype: string
    - name: context_id
      dtype: int32
    - name: question_id
      dtype: int32
    - name: symbolic_context
      dtype: string
    - name: symbolic_entity_map
      dtype: string
    - name: symbolic_question
      sequence: string
    - name: symbolic_reasoning
      dtype: string
    - name: num_context_entities
      dtype: int32
    - name: num_question_entities
      dtype: int32
    - name: question_type
      dtype: string
    - name: reasoning_types
      sequence: string
    - name: spatial_types
      sequence: string
    - name: commonsense_question
      dtype: string
    - name: canary
      dtype: string
    - name: comments
      sequence: string
configs:
  - config_name: SpaRP-PS1 (SpaRTUN)
    version: 1.1.0
    data_files:
      - split: train
        path: sparp/SpaRP-PS1 (SpaRTUN)/train.json
      - split: validation
        path: sparp/SpaRP-PS1 (SpaRTUN)/val.json
      - split: test
        path: sparp/SpaRP-PS1 (SpaRTUN)/test.json
  - config_name: SpaRP-PS2 (StepGame)
    version: 1.1.0
    data_files:
      - split: train
        path: sparp/SpaRP (StepGame)/PS2/train.json
      - split: validation
        path: sparp/SpaRP (StepGame)/PS2/val.json
      - split: test
        path: sparp/SpaRP (StepGame)/PS2/test.json
  - config_name: SpaRP-PS3 (StepGame-Ext-01)
    version: 1.1.0
    data_files:
      - split: train
        path: sparp/SpaRP (StepGame)/PS3/train.json
      - split: validation
        path: sparp/SpaRP (StepGame)/PS3/val.json
      - split: test
        path: sparp/SpaRP (StepGame)/PS3/test.json
  - config_name: SpaRP-PS4 (StepGame-Ext-02)
    version: 1.1.0
    data_files:
      - split: train
        path: sparp/SpaRP (StepGame)/PS4/train.json
      - split: validation
        path: sparp/SpaRP (StepGame)/PS4/val.json
      - split: test
        path: sparp/SpaRP (StepGame)/PS4/test.json
  - config_name: small-SpaRP-PS1 (SpaRTUN)
    version: 1.1.0
    data_files:
      - split: train
        path: small-sparp/SpaRP-PS1 (SpaRTUN)/train.json
      - split: validation
        path: small-sparp/SpaRP-PS1 (SpaRTUN)/validation.json
      - split: test
        path: small-sparp/SpaRP-PS1 (SpaRTUN)/test.json
  - config_name: small-SpaRP-PS2 (StepGame)
    version: 1.1.0
    data_files:
      - split: train
        path: small-sparp/SpaRP (StepGame)/PS2/train.json
      - split: validation
        path: small-sparp/SpaRP (StepGame)/PS2/validation.json
      - split: test
        path: small-sparp/SpaRP (StepGame)/PS2/test.json
  - config_name: small-SpaRP-PS3 (StepGame-Ext-01)
    version: 1.1.0
    data_files:
      - split: train
        path: small-sparp/SpaRP (StepGame)/PS3/train.json
      - split: validation
        path: small-sparp/SpaRP (StepGame)/PS3/validation.json
      - split: test
        path: small-sparp/SpaRP (StepGame)/PS3/test.json
  - config_name: small-SpaRP-PS4 (StepGame-Ext-02)
    version: 1.1.0
    data_files:
      - split: train
        path: small-sparp/SpaRP (StepGame)/PS4/train.json
      - split: validation
        path: small-sparp/SpaRP (StepGame)/PS4/validation.json
      - split: test
        path: small-sparp/SpaRP (StepGame)/PS4/test.json

Dataset Card for Spatial Reasoning Path (SpaRP)

Table of Contents
Dataset Description
- Dataset Summary
- Languages
Additional Information

Dataset Description

Repository: https://github.com/UKPLab/acl2024-sparc-and-sparp
Paper: https://arxiv.org/abs/
Point of Contact: Md Imbesat Hassan Rizvi (http://www.ukp.tu-darmstadt.de/)

Dataset Summary

This dataset is a consolidation of SpaRTUN and StepGame datasets with an extension of additional spatial characterization and reasoning path generation. The methodology is explained in our ACL 2024 paper - SpaRC and SpaRP: Spatial Reasoning Characterization and Path Generation for Understanding Spatial Reasoning Capability of Large Language Models. The dataset format and fields are normalized across the two upstream benchmark datasets -- SpaRTUN and StepGame. The datasets are primarily a Spatial Question Answering datasets, which are enriched with verbalized reasoning paths. The prominent fields of interests are:

context: Textual description of the spatial context.
question: A question about finding spatial relations between two entities in the context.
targets: Answer i.e. list of spatial relations between the entities in the question.
target_choices: List of all the spatial relations to choose from.
target_scores: Binarized Multi-label representation of targets over target_choices.
reasoning: Verbalized reasoning path as deductively-verified CoT for training or few-shot examples.

Additionally, the fields with the metadata are:

context_id: An identifier from the source data corresponding to the context. context_id is unique. A single context can have multiple questions (e.g. SpaRTUN). Hence, (context_id, question_id) is a unique identifier for a dataset instance.
question_id: An identifier from the source data corresponding to the question.
num_hop: Ground truth number of hop required for the question.
symbolic_context: A json string describing the symbolic context.
symbolic_entity_map: A json string that maps symbolic entities to their complete descriptive names used.
symbolic_question: A list containing head and tail entities of the question.
symbolic_reasoning: A json string containing symbolic reasoning steps.
num_context_entities: Number of entities in the context.
num_question_entities: Number of entities in the question.
question_type: Type of the question. Only FR i.e. Find Relation type questions are currently present.
canary: A canary string present only in the test.
reasoning_types: The type of reasoning copied from the source data required for answering the question.
spatial_types: The type of spatial relations copied from the source data required for answering the question.
source_data: The upstream source of the data (either SpaRTUN or StepGame) for a given instance.
comments: Additional comments specific to the upstream data.

Languages

English

Additional Information

You can download the data via:

from datasets import load_dataset

dataset = load_dataset("UKPLab/sparp") # default config is "SpaRP-PS1 (SpaRTUN)"
dataset = load_dataset("UKPLab/sparp", "SpaRP-PS2 (StepGame)") # use the "SpaRP-PS2 (StepGame)" tag for the StepGame dataset

Please find more information about the code and how the data was collected on GitHub.

Dataset Curators

Curation is managed by our data manager at UKP.

Licensing Information

CC-by-SA 4.0

Citation Information

Please cite this data using:

@inproceedings{rizvi-2024-sparc,
  title={SpaRC and SpaRP: Spatial Reasoning Characterization and Path Generation for Understanding Spatial Reasoning Capability of Large Language Models},
  author={Rizvi, Md Imbesat Hassan Rizvi and Zhu, Xiaodan and Gurevych, Iryna},
  editor = "",
  booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics",
  month = aug,
  year = "2024",
  address = "Bangkok, Thailand",
  publisher = "Association for Computational Linguistics",
  url = "",
  doi = "",
  pages = "",
}