Explore Kinship.
Follow the evidence.

Browse real Kinship questions

Explore the published questions, reference answers, and complete entity-relation paths without downloading files or running Python. The labels are already human-readable, so no Wikidata or label service is needed.

Load the dataset in Python

Use huggingface_hub to download the repository snapshot, then load the graph and questions into pandas DataFrames.

python -m pip install huggingface_hub pandas

Choose a QA file

from pathlib import Path
import pandas as pd
from huggingface_hub import snapshot_download

# Download the Kinship dataset from Hugging Face.
root = Path(snapshot_download(
    repo_id="HalcyonSolutions/Kinship",
    repo_type="dataset",
))

# The original graph is used for navigation.
triples = pd.read_csv(
    root / "kg/orig/triplets.txt",
    sep="\t",
    header=None,
    names=["head", "relation", "tail"],
)

# Load natural-language questions and gold answers.
qa = pd.read_csv(root / "qa/kinship_qa_nhop.csv")

print(triples.head())
print(qa[["Question", "Source-Entity", "Answer-Entity", "Hops"]].head())

What's included?

The repository contains graph triples, QA records with reference paths, and the original Kinship source data. For navigation experiments, use kg/orig/.

Read the annotated reasoning

Every QA record has an explicit source entity and gold path. Paths are stored as string representations of ordered triples in the Paths column.

import ast

# Decode one gold reasoning path.
example = qa.loc[qa["Hops"] == 2].iloc[0]
path = ast.literal_eval(example["Paths"])

print("Question:", example["Question"])
print("Answer:", example["Answer"])
for head, relation, tail in path:
    print(f"{head} --[{relation}]--> {tail}")

# QA split labels are stored in the CSV itself.
test_questions = qa.loc[qa["SplitLabel"] == "test"]
print("Test questions:", len(test_questions))

The example above reads ground-truth annotations. It does not run MINERVA or generate predictions. The QA records use human-readable entity labels.

Next steps

For navigation experiments, pretrained checkpoints, and methods, follow the companion resources.

License: CC BY 4.0. If you use Kinship, please credit the original Hinton (1986) dataset and the THESEUS work as described on the dataset card.