naos-ku/motif-qa

MotifQA

Dataset Summary

MotifQA is a synthetic graph question-answering benchmark focused on detecting graph motifs inside small random graphs. Each example pairs a textual prompt with an answer sentence, a list of nodes highlighted as the motif (when present), and an explicit graph description(nodes and edges). In this QA dataset, all graphs are homogenous and undirected. Subsets cover both yes/no motif detection and motif-type classification (house vs 5-cycle).

Subsets

• ba_shapes (default): Detect a five-node house motif inside Barabási–Albert graphs.
• tree_cycle: Detect a 6-node cycle inside random binary trees.
• tree_grid: Detect a 3×3 grid motif inside random binary trees.
• ba_two_motifs: Classify whether a Barabási–Albert graph contains a house motif or a 5-cycle motif (every graph has exactly one of the two).

Dataset Structure

Data Fields

• prompt (string): Natural-language question about the presence of motif.
• response (string): Short-form answer sentence. e.g., "Yes, there is a motif." or "No, there is no motif.".
• motif_nodes (List[int]): Node indices comprising the motif; empty list when the motif is absent.
• nodes (List[int]): Enumeration of all node identifiers in the graph (ranges from 0 to nnodes - 1).
• edges (List[List[int]]): List of undirected edges; every edge appears once as [min_node, max_node].
• nnodes (int): Total number of nodes.
• nedges (int): Total number of edges.

Data Splits

Subset	Split	Total	Positive	Negative
ba_shapes	Train	1000	511	489
ba_shapes	Val	500	246	254
ba_shapes	Test	500	243	257
tree_cycle	Train	1000	496	504
tree_cycle	Val	500	245	255
tree_cycle	Test	500	259	241
tree_grid	Train	1000	508	492
tree_grid	Val	500	242	258
tree_grid	Test	500	250	250

Detection subsets are approximately balanced across splits. The ba_two_motifs subset contains only motif-bearing graphs; label counts are:

Subset	Split	Total	House	Cycle
ba_two_motifs	Train	1000	513	487
ba_two_motifs	Val	500	241	259
ba_two_motifs	Test	500	246	254

Node Count Distribution

Node counts depend on the underlying base generator and motif size:

• ba_shapes (Barabási–Albert, num_edges = 2): 8–20 nodes.
• tree_cycle (Prüfer trees + 6-cycle motif): 8–21 nodes.
• tree_grid (Prüfer trees + 3×3 grid motif): 8–24 nodes.
• ba_two_motifs (Barabási–Albert; house vs 5-cycle motifs): 8–20 nodes.

Observed frequencies across the full 2,000-example splits for each subset are shown in the per-subset histograms.

Data Creation

Source Data

All examples are synthetically generated using PyTorch Geometric (2.7.0) primitives:

• Detection subsets: positive samples start from certain base graphs whose node counts are drawn from a uniform distribution before a motif is injected, yielding final graphs with 8–20 nodes. Negative samples use the same process with node counts drawn from range(MIN_NODES, MAX_NODES + 1) (default 8–20) and no motif.
• ba_two_motifs: both classes contain motifs injected into Barabási–Albert graphs—house motifs vs 5-cycle motifs—with final graphs spanning 8–20 nodes.

Generation Process

1. Create class-specific graph pools (ExplainerDataset for motif-bearing graphs and NegativeGraphs for motif-free graphs, or two different ExplainerDataset instances for ba_two_motifs).
2. Randomly shuffle node indices via the custom ShuffleNodes transform to prevent positional bias.
3. Convert graphs to text QA format:
   • Construct the prompt and answer sentence.
   • Record motif node indices and extract the explicit graph topology.
4. Shuffle and split into train/validation/test partitions (50/25/25).
5. Serialize each subset to JSON Lines (train.jsonl, val.jsonl, test.jsonl).

Visualization utilities (src/utils.py) can render example graphs with highlighted motifs; pre-generated figures are stored in fig/.

Annotations

Annotations are automatically derived from the generation pipeline; no human annotation was performed.

Considerations for Using the Data

• Edges are stored once per undirected connection with node identifiers sorted; duplicate each entry if your pipeline expects bidirectional edges.
• Natural-language responses are templated sentences, so linguistic diversity is limited.

Usage

Loading with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("naoki/motif-qa", name="ba_shapes")
print(ds["train"][0]["response"])

Working with Graph Topology

import torch
from torch_geometric.data import Data
from src.utils import visualize_graph

example = ds["train"][0]

# Build a PyG data object from the serialized edge list.
edge_index = torch.tensor(example["edges"], dtype=torch.long).t().contiguous()
graph = Data(edge_index=edge_index, num_nodes=example["nnodes"])

# Attach the motif mask expected by `visualize_graph`.
graph.node_mask = torch.zeros(example["nnodes"])
graph.node_mask[example["motif_nodes"]] = 1

visualize_graph(graph, "motif_example.png")

The script saves a PNG with motif nodes highlighted in orange. Use absolute imports (e.g., from motif_qa.src.utils import visualize_graph) if you install the project as a package.

Regenerating the Dataset Locally

python generator.py \
  --num-pos-samples 1000 \
  --num-neg-samples 1000 \
  --out-dir data

This command reproduces 2,000 examples and overwrites the JSONL files.

Validating with datasets-cli

Run the standard Hugging Face check from the repository root:

datasets-cli test . --all_configs

The command materializes the ba_shapes subset via the built-in JSON loader.

Limitations and Biases

• Only a house motif is considered; absence of other motifs may bias models toward motif-specific heuristics.
• Graph topologies follow Barabási–Albert statistics; performance on other graph families is untested.
• Answers are deterministic and class-balanced, so real-world noise is not represented.

License

Released under the MIT License.

Citation

If you use this dataset, please cite the repository:

@software{motifqa2025,
  author = {Naoki Shimoda},
  title = {MotifQA: Synthetic Graph Motif Question Answering Dataset},
  year = {2025},
  url = {https://huggingface.co/datasets/naos-ku/motif-qa}
}