zhougengxian/CoG-Evaluation-Data

CoG Evaluation Data

The seven evaluation sets used in CoG (EMNLP 2026), comprising 5,174 questions for knowledge-intensive QA across KG-based and text-based benchmarks. This collection supports evaluation of multi-hop retrieval and reasoning with graph, text, and hybrid RAG methods.

Paper · Code

Datasets

Dataset / HF config	Knowledge source	Questions	Raw file
KGQAGen / kgqagen	KG	1,079	KGQAGen-10k.json
CWQ / cwq	KG	1,024	cwq.json
QALD10-en / qald10_en	KG	333	qald_10-en.json
WebQSP / webqsp	KG	1,430	webqsp_test.json
2WikiMultiHopQA / 2wikimultihopqa	Text	500	2wikimultihopqa.json
AdvHotpotQA / advhotpotqa	Text	308	hotpotadv_dev.json
MuSiQue / musique	Text	500	musique.json

Counts and KG/Text categories follow the CoG paper. All configurations use test to denote evaluation use. Original files are under raw/; retrieval corpora and indexes are configured separately through the codebase.

Loading

Load KGQAGen from Hugging Face:

import json
from datasets import load_dataset

ds = load_dataset("zhougengxian/CoG-Evaluation-Data", "kgqagen", split="test")
example = ds[0]
question = example["question"]
answers = json.loads(example["answer"])

In the HF Parquet files, list and dictionary values are JSON-encoded strings; decode them with json.loads. Scalar strings remain unchanged.

Raw JSON file

Download the original JSON structure for use with the CoG code:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "zhougengxian/CoG-Evaluation-Data",
    "raw/KGQAGen-10k.json",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as f:
    records = json.load(f)
question = records[0]["question"]
answers = records[0]["answer"]

Key fields

The table lists fields for questions, reference answers, and question entities. Answer types refer to the raw JSON files.

HF config	Question field	Answer field (raw type)	Question entity field
kgqagen	question	answer (list)	entities (separate annotation file)
cwq	question	answer (string)	qid_topic_entity
qald10_en	question	answer (dictionary)	qid_topic_entity
webqsp	question	answers (list)	qid_topic_entity
2wikimultihopqa	question	answer (list)	entities (separate annotation file)
advhotpotqa	question	answer (string)	entities (separate annotation file)
musique	question	answer (list)	entities (separate annotation file)

Question entity annotations

Entity annotations provide seed nodes for models that require starting entities for graph retrieval. For KGQAGen, 2WikiMultiHopQA, AdvHotpotQA, and MuSiQue, respectively, the files under raw/ are KGQAGen-10k_entities_azure.json, 2wikimultihopqa_entities_azure.json, hotpotadv_entities_azure.json, and musique_entities_azure.json. Load their HF versions by appending _entities to the primary configuration name.

For example, load the KGQAGen question entity annotations:

import json
from datasets import load_dataset

annotations = load_dataset(
    "zhougengxian/CoG-Evaluation-Data",
    "kgqagen_entities",
    split="test",
)
question = annotations[0]["question"]
entities = json.loads(annotations[0]["entities"])

These files contain question and entities; match the question to the primary file for reference answers. CWQ, QALD10-en, and WebQSP already include qid_topic_entity. The default CoG pipeline does not require supplied annotations.

Attribution and licensing

See SOURCES.md.

Citation

@article{zhou2026cog,
  title={Cognition on Graph: Navigating Massive Knowledge Space via Cognitive Cycles and Bidirectional Graph-Text Synergy}, 
  author={Gengxian Zhou and Jian Xu and Zichen Tang and Shiming Xiang and Haihong E and Cheng-Lin Liu},
  year={2026},
  eprint={2609.12791},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2609.12791}, 
}

Also cite the original benchmark publications when using their records.