Datasetpublic

Allen-UQ/CNY-data

CNY walk-ready graph reasoning data Walk-ready data for CNY (Call Neighbours Yourself), the graph-walk reinforcement learning framework described in Call Neighbours Yourself: Graph Walks with Destination-Conditioned On-Policy Self-Distillation (EMNLP 2026), arXiv:2608.29588. Models: Allen-UQ/CNY-7B, Allen-UQ/CNY-14B. This repository does not redistribute any graph dataset in its original form. It provides the derived, prompt-rendered form in which each graph instance is… See the full description on the dataset page: https://huggingface.co/datasets/Allen-UQ/CNY-data.

sourceHugging Faceotherupdated 24d agoView on Hugging Face

0likes435downloads

Dataset Card

CNY walk-ready graph reasoning data

![arXiv](https://arxiv.org/abs/2608.29588) ![GitHub](https://github.com/superallen13/CNY) ![Model](https://huggingface.co/Allen-UQ/CNY-14B)

Walk-ready data for CNY (Call Neighbours Yourself), the graph-walk reinforcement learning framework described in Call Neighbours Yourself: Graph Walks with Destination-Conditioned On-Policy Self-Distillation (EMNLP 2026), arXiv:2608.29588.

Models: `Allen-UQ/CNY-7B`, `Allen-UQ/CNY-14B`.

This repository does not redistribute any graph dataset in its original form. It provides the derived, prompt-rendered form in which each graph instance is presented to the walk policy: an ego-node text, the label descriptions and one short preview line per neighbour, with a neighbour's full text withheld until the model emits a <walk> action.

Format

One JSON object per line, in the schema the training stack consumes directly (input_key = messages, label_key = solution):

field	type	content
`messages`	list	chat turns; the user turn carries the rendered walk prompt
`solution`	str	JSON string with `ground_truth`, `dataset_name`, `task_type`
`metadata`	dict	reserved, empty in this release

python

from datasets import load_dataset

ds = load_dataset("Allen-UQ/CNY-data", "cora-7way", split="test")

Evaluation configs

Every directory carries the config_snapshot.yaml that generated it, so the exact preview length, neighbour cap, topology radius and random state are recoverable.

config	rows	n-way	source cache version
`cora-7way`	540	all 7	`cora/06fc44e652981591`
`cora-2way`	540	2	`cora/9025a930a351fc1d`
`wikics-10way`	2,340	all 10	`wikics/d11d65bf1f380875`
`wikics-5way`	2,340	5	`wikics/d0e6c4699e60da6e`
`products-10way`	3,000	10	`products_sub/8f7441753fd26449`
`products-5way`	3,000	5	`products_sub/9cde029277a0b020`
`products-full`	37,745	n-way sampled at evaluation	`products/2c35a69d89f14c18`
`fb15k237-10way`	20,466	10	`fb15k237_edge_tag/3a282e77d408bd39`
`expla-graph`	554	2 stances	`expla_graph_walk/640b55c2e57ccda6`
`webqsp`	1,628	open-ended	`webqsp_walk/216e2599b263ef3a`

fb15k237-10way and expla-graph are the exact versions behind the corresponding Table 1 entries of the paper. The Cora, WikiCS and Products configs are the versions used by the paired multi-seed evaluation runs; products-full matches the 37,745-row test split reported in the paper's dataset statistics.

Graphs

raw_datasets/<dataset>.pt holds the graphs a <walk> traverses. A prompt only carries the ego-node text and neighbour previews, so the graph is required to reveal a destination's full text during rollout.

file	role
`citeseer` `pubmed` `photo` `computer` `history` `sportsfit` `instagram` `wn18rr_edge_tag`	training graphs
`cora` `wikics` `products` `products_sub` `fb15k237_edge_tag` `expla_graph_walk` `webqsp_walk`	evaluation graphs
`wikics_shuffle`	rewired control used by the topology-versus-semantics analysis
`products_sub_test_indices.json`	index defining the fixed Products subsets

Training configs

train is the mixture the reported models were trained on. The other two expose its two halves for analysis.

config	rows	content
`train`	317,495	the full 8-dataset mixture actually used for training
`train-node-mixture`	230,660	its node-classification half, seven graphs
`train-wn18rr-edge`	86,835	its relation-classification half, WN18RR

Composition of train: node_class 230,660 (Sportsfit 103,834, Computer 52,336, Photo 29,016, History 24,930, PubMed 11,830, Instagram 6,803, CiteSeer 1,911) and relation_class 86,835 (WN18RR). The totals match the training splits in the paper's dataset statistics table.

Provenance and licensing

Each config is a derivative of a published graph dataset. Copyright in the underlying node text remains with the original owners, and use of a config is governed by the licence of its source.

config group	source dataset	original release	licence of source
`cora-*`	Cora	Sen et al., 2008	academic use
`wikics-*`	WikiCS	Mernyei and Cangea, 2020	derived from Wikipedia, CC BY-SA 3.0
`products-*`	ogbn-products	Hu et al., 2020 (OGB)	OGB code MIT; underlying Amazon product data subject to its own terms
`fb15k237-10way`	FB15k-237	Toutanova and Chen, 2015	derived from Freebase, CC BY 2.5
`expla-graph`	ExplaGraphs	Saha et al., 2021	see original release
`webqsp`	WebQSP	Yih et al., 2016	see original release
`train-node-mixture`	CiteSeer, PubMed, and the Amazon-derived Photo, Computer, History, Sportsfit, plus Instagram	respective authors	mixed; the Amazon-derived subsets carry the same terms as `products-*`
`train-wn18rr-edge`	WN18RR	Dettmers et al., 2018	derived from WordNet, WordNet licence

Users must satisfy the terms of the original sources. Anyone requiring a distribution free of third-party text can instead rebuild every config from source: the random state is fixed at 42 and construction is deterministic, so the config_snapshot.yaml in each directory reproduces the same splits, neighbourhoods and previews.

Citation

bibtex

@inproceedings{liu2026cny,
  title     = {Call Neighbours Yourself: Graph Walks with Destination-Conditioned On-Policy Self-Distillation},
  author    = {Yilun Liu and Boyu Luo and Yanran Tang and Ruihong Qiu and Zi Huang},
  booktitle = {EMNLP},
  year      = {2026}
}

Please also cite the source dataset of any config used.