license: apache-2.0
configs:
- config_name: dataset_da_huse_2x3_5rh
data_files:
- split: train
path: dataset_da_huse_2x3_5rh/train-*
- split: test
path: dataset_da_huse_2x3_5rh/test-*
- config_name: dataset_da_huse_4x5_5rh
data_files:
- split: train
path: dataset_da_huse_4x5_5rh/train-*
- split: test
path: dataset_da_huse_4x5_5rh/test-*
- config_name: dataset_da_smoerrebroed_2x3_5rh
data_files:
- split: train
path: dataset_da_smoerrebroed_2x3_5rh/train-*
- split: test
path: dataset_da_smoerrebroed_2x3_5rh/test-*
- config_name: dataset_da_smoerrebroed_4x5_5rh
data_files:
- split: train
path: dataset_da_smoerrebroed_4x5_5rh/train-*
- split: test
path: dataset_da_smoerrebroed_4x5_5rh/test-*
- config_name: dataset_de_hauser_2x3_5rh
data_files:
- split: train
path: dataset_de_hauser_2x3_5rh/train-*
- split: test
path: dataset_de_hauser_2x3_5rh/test-*
- config_name: dataset_de_hauser_4x5_5rh
data_files:
- split: train
path: dataset_de_hauser_4x5_5rh/train-*
- split: test
path: dataset_de_hauser_4x5_5rh/test-*
- config_name: dataset_en_houses_2x3_5rh
data_files:
- split: train
path: dataset_en_houses_2x3_5rh/train-*
- split: test
path: dataset_en_houses_2x3_5rh/test-*
- config_name: dataset_en_houses_4x5_5rh
data_files:
- split: train
path: dataset_en_houses_4x5_5rh/train-*
- split: test
path: dataset_en_houses_4x5_5rh/test-*
- config_name: dataset_fo_hus_2x3_5rh
data_files:
- split: train
path: dataset_fo_hus_2x3_5rh/train-*
- split: test
path: dataset_fo_hus_2x3_5rh/test-*
- config_name: dataset_fo_hus_4x5_5rh
data_files:
- split: train
path: dataset_fo_hus_4x5_5rh/train-*
- split: test
path: dataset_fo_hus_4x5_5rh/test-*
- config_name: dataset_is_husum_2x3_5rh
data_files:
- split: train
path: dataset_is_husum_2x3_5rh/train-*
- split: test
path: dataset_is_husum_2x3_5rh/test-*
- config_name: dataset_is_husum_4x5_5rh
data_files:
- split: train
path: dataset_is_husum_4x5_5rh/train-*
- split: test
path: dataset_is_husum_4x5_5rh/test-*
- config_name: dataset_nb_hus_2x3_5rh
data_files:
- split: train
path: dataset_nb_hus_2x3_5rh/train-*
- split: test
path: dataset_nb_hus_2x3_5rh/test-*
- config_name: dataset_nb_hus_4x5_5rh
data_files:
- split: train
path: dataset_nb_hus_4x5_5rh/train-*
- split: test
path: dataset_nb_hus_4x5_5rh/test-*
- config_name: dataset_nn_hus_2x3_5rh
data_files:
- split: train
path: dataset_nn_hus_2x3_5rh/train-*
- split: test
path: dataset_nn_hus_2x3_5rh/test-*
- config_name: dataset_nn_hus_4x5_5rh
data_files:
- split: train
path: dataset_nn_hus_4x5_5rh/train-*
- split: test
path: dataset_nn_hus_4x5_5rh/test-*
- config_name: dataset_sv_hus_2x3_5rh
data_files:
- split: train
path: dataset_sv_hus_2x3_5rh/train-*
- split: test
path: dataset_sv_hus_2x3_5rh/test-*
- config_name: dataset_sv_hus_4x5_5rh
data_files:
- split: train
path: dataset_sv_hus_4x5_5rh/train-*
- split: test
path: dataset_sv_hus_4x5_5rh/test-*
dataset_info:
- config_name: dataset_da_huse_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 186251
num_examples: 128
- name: test
num_bytes: 1486529
num_examples: 1024
download_size: 203726
dataset_size: 1672780
- config_name: dataset_da_huse_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 386741
num_examples: 128
- name: test
num_bytes: 3057562
num_examples: 1024
download_size: 512397
dataset_size: 3444303
- config_name: dataset_da_smoerrebroed_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 226795
num_examples: 128
- name: test
num_bytes: 1806513
num_examples: 1024
download_size: 219013
dataset_size: 2033308
- config_name: dataset_da_smoerrebroed_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 445587
num_examples: 128
- name: test
num_bytes: 3571817
num_examples: 1024
download_size: 547859
dataset_size: 4017404
- config_name: dataset_de_hauser_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 210461
num_examples: 128
- name: test
num_bytes: 1681221
num_examples: 1024
download_size: 212285
dataset_size: 1891682
- config_name: dataset_de_hauser_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 414711
num_examples: 128
- name: test
num_bytes: 3306651
num_examples: 1024
download_size: 525426
dataset_size: 3721362
- config_name: dataset_en_houses_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 187565
num_examples: 128
- name: test
num_bytes: 1498622
num_examples: 1024
download_size: 200675
dataset_size: 1686187
- config_name: dataset_en_houses_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 390474
num_examples: 128
- name: test
num_bytes: 3130712
num_examples: 1024
download_size: 513798
dataset_size: 3521186
- config_name: dataset_fo_hus_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 206735
num_examples: 128
- name: test
num_bytes: 1647768
num_examples: 1024
download_size: 216372
dataset_size: 1854503
- config_name: dataset_fo_hus_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 419235
num_examples: 128
- name: test
num_bytes: 3352975
num_examples: 1024
download_size: 540386
dataset_size: 3772210
- config_name: dataset_is_husum_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 196218
num_examples: 128
- name: test
num_bytes: 1579934
num_examples: 1024
download_size: 204625
dataset_size: 1776152
- config_name: dataset_is_husum_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 416366
num_examples: 128
- name: test
num_bytes: 3329711
num_examples: 1024
download_size: 532681
dataset_size: 3746077
- config_name: dataset_nb_hus_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 183889
num_examples: 128
- name: test
num_bytes: 1464132
num_examples: 1024
download_size: 200028
dataset_size: 1648021
- config_name: dataset_nb_hus_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 376601
num_examples: 128
- name: test
num_bytes: 3022073
num_examples: 1024
download_size: 506897
dataset_size: 3398674
- config_name: dataset_nn_hus_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 183493
num_examples: 128
- name: test
num_bytes: 1478353
num_examples: 1024
download_size: 200746
dataset_size: 1661846
- config_name: dataset_nn_hus_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 379289
num_examples: 128
- name: test
num_bytes: 3036246
num_examples: 1024
download_size: 510417
dataset_size: 3415535
- config_name: dataset_sv_hus_2x3_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 186037
num_examples: 128
- name: test
num_bytes: 1477385
num_examples: 1024
download_size: 203972
dataset_size: 1663422
- config_name: dataset_sv_hus_4x5_5rh
features:
- name: introduction
dtype: string
- name: clues
sequence: string
- name: question
dtype: string
- name: format_instructions
dtype: string
- name: format_example
dtype: string
- name: solution
struct:
- name: object_1
sequence: string
- name: object_2
sequence: string
- name: object_3
sequence: string
- name: object_4
sequence: string
- name: clue_types
sequence: string
- name: red_herrings
sequence: int64
splits:
- name: train
num_bytes: 386173
num_examples: 128
- name: test
num_bytes: 3078935
num_examples: 1024
download_size: 519763
dataset_size: 3465108
Dataset Card for Dataset Name
This dataset includes zebra puzzles in multiple European languages and in two sizes: 2x3 and 4x5. It can be used for evaluating logical reasoning ability.
The data has been generated using the code in this repo.
Dataset Details
Dataset Description
Zebra puzzles are a type of constraint satisfaction problem. They describe a number of objects, N_objects, that each have attributes, N_attributes. The goal is to couple the objects with the correct attributes, given some clues. Each solution can be described as a N_objects x N_attributes matrix.
To increase difficulty, we include "red herrings" which follow the same structure as true clues, but contain no relevant information. We use 5 red herrings per puzzle.
Most dataset folders contain puzzles with the "houses" theme, where the objects are houses and each attribute describes an inhabitant. Attributes are randomly selected categories such as nationalities and jobs. This is included in Danish, English, Faroese, German, Icelandic, Norwegian Bokmål, Norwegian Nynorsk and Swedish. A Dutch version is in progress.
We are currently testing changes to template phrasing in the Danish version, so these may differ slightly from the rest.
We also include data with the "smøerrebrød" theme, where the objects are smørrebrød (open sandwiches) and each attribute is an ingredient. Categories are ingredient types such as bread or garnish. This theme is only included in Danish.
The dataset includes puzzles, solutions, lists of included clue types and indices of red herring clues.
The training sets contain 128 puzzles each which are meant as examples for practise. The test sets contain 1024 puzzles each.
- Curated by: Sofie Helene Bruun (sofie.bruun@alexandra.dk) from the Alexandra Institute.
- Funded by [optional]: [More Information Needed]
- Language(s) (NLP): Danish (da), English (en), Faroese (fo), German (de), Icelandic (is), Norwegian Bokmål (nb), Norwegian Nynorsk (nn) and Swedish (sv).
- License: apache-2.0
Dataset Sources
- Repository: https://github.com/alexandrainst/zebra_puzzles
- Paper: In progress.
Uses
Logical reasoning ability can be evaluated by comparing reponses to puzzles to the true solutions. For examples of how this can be done, see the associated repository.
The dataset contains examples of the suggested JSON format of reponses for evaluation of LLM's.
Part of the dataset is intended for use in EuroEval.
Direct Use
Each puzzle can be combined from the columns: introduction, clues and question.
Evaluation can be performed by comparing a response to the solution column.
When evaluating LLM's, consider including the format_instructions and format_example columns in each puzzle, so it is clear how the intended response should be formatted.
To create puzzles without red herrings, remove the clues with indices defined in the red_herrings column.
To analyse the effect of different clue types and red herrings, the clue_types and red_herrings columns can be compared to model performance.
Out-of-Scope Use
The clue_types and red_herrings columns should not be included during the solving process, as they will reduce the need for understanding the natural language prompt. Of course, the solution column should also not be included.
Dataset Structure
Each puzzle is generated randomly and independently of other puzzles.
The columns are:
- introduction (str): Defines the overall rules and introduces the attributes.
- clues (list[str]): Clues relating the attributes and objects. 5 red herrings are included per puzzle.
- question (str): The question to answer by a solution.
- format_instructions (str): Instructions on how to respond in JSON format. This is relevant for LLM's.
- format_example (str). An example of the reponse format with the included attribute categories from the puzzle (but not the exact attributes).
- solution. The solution matrix in JSON format.
- clue_types. The list of clue types matching the clues column.
- red_herrings. A list of indices to the red herring clues.
Dataset Creation
Curation Rationale
The motivation is creating a multilingual benchmark for logical reasoning. The data allows us to compare logical reasoning ability of LLM's and compare scores across languages.
Most of the dataset follows the traditional house theme, which is easy to translate. The smørrebrød theme is included to make it possible to compare the house theme to puzzles matching a European culture tied to a specific language.
Source Data
The data is created from words and phrases defined in the zebra puzzle repository.
Data Collection and Processing
The included words and phrases have been drafted by the author with the help of Google Translate, GPT-4.1 in Github Copilot, dictionaries and Wikipedia. Relevant code and a few puzzles have been evaluated by native/fluent speakers of each included language. More details are included in the paper in progress.
Who are the source data producers?
Sofie Helene Bruun from the Alexandra Institute with help from other people involved in LLM evaluation across Europe.
Annotations [optional]
Annotation process
The puzzles and solutions are automatically generated to match.
Who are the annotators?
None.
Personal and Sensitive Information
No personal or sensitive information is included.
Bias, Risks, and Limitations
Not every combination of words in the dataset has been read by a native speaker of each language, so there is a risk that an included combination sounds unnatural or creates an unintended meaning.
Attributes are combined randomly and might accidentally match stereotypes or traits of real people.
The randomly generated smørrebrød are typically not representative of traditional Danish cuisine, although many of the ingredients are.
Recommendations
Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
Citation [optional]
If you use this dataset in your research, please cite our paper:
BibTeX:
[Paper in progress]
APA:
[Paper in progress]