Thank you for your interest in CAPC-CG — we aim to review requests within a week. CAPC-CG is released under CC BY-NC 4.0 for non-commercial research use only.

📝 A few tips to help us review smoothly:

Please use your institutional email. If only a personal email is available, include a link to your ORCID, Google Scholar, or institutional profile.
Provide your full name and institution as they appear in your publications (e.g., "Tsinghua University" rather than "thu").
For "Research Purpose", 2–3 sentences is ideal — single-word answers like "research" are hard to evaluate.

All submissions are logged. Access may be revoked if information is later found to be inaccurate or if the data is used in ways inconsistent with our license terms.

By submitting this form, you agree that:

(1) you will use CAPC-CG solely for non-commercial academic or research purposes;
(2) you will cite the CAPC-CG paper (ACL 2026) in any publication, presentation, blog post, or other public artifact resulting from use of this dataset;
(3) you will not redistribute any portion of the data, in whole or in part, to third parties;
(4) you will comply with all applicable laws and institutional ethics requirements.

Misuse or false information may result in revocation of access.

CAPC-CG: A Large-Scale, Expert-Directed LLM-Annotated Corpus of Adaptive Policy Communication in China

🎉 Accepted to ACL 2026 (Main Conference) — if you use this dataset in any form, please be sure to cite our paper (see Citation below).

CAPC-CG is the first large-scale open dataset of Chinese central-government policy directives (1949–2023), annotated with a theory-based five-color typology of policy signals — Black (Authorizing), Yellow (Pressuring), Charcoal (Flexible), Grey (Ambiguous), Red (Prohibiting) — grounded in Ang's theory of adaptive policy communication. The corpus covers national laws, administrative regulations, and ministerial rules issued by China's top authorities, segmented into ≈3.3 million paragraph-level units.

Alongside the full corpus, we release a gold-standard annotated subset used for training and evaluating classifiers, with a reported inter-annotator agreement of Fleiss' κ = 0.86 on directive labels. Baseline models (SVM, XGBoost, BERT-base-Chinese, Qwen2.5-7B, Llama-3-8B, GPT-4o-mini zero/few-shot/fine-tuned) and the full annotation codebook are distributed alongside the data.

Authors: Bolun Sun¹·², Charles Chang³, Yuen Yuen Ang¹, Ruotong Mu¹, Yuchen Xu¹, Zhengxin Zhang¹, Pingxu Hao¹ ¹ Johns Hopkins University · ² Northwestern University · ³ Duke Kunshan University

🗂 Dataset Structure

File	Rows	Size	Description
`data/CN_DOCS_FINAL_complete.parquet`	3,275,474	~500 MB	Full segmented corpus with metadata & auto-labels
`task1_level1/train.parquet`	600	100 KB	Level-1 gold-standard train split (chat format, preserved as `messages` column)
`task1_level1/validation.parquet`	300	55 KB	Level-1 gold-standard val split
`task2_level2/train.parquet`	1,200	265 KB	Level-2 gold-standard train split
`task2_level2/validation.parquet`	400	88 KB	Level-2 gold-standard val split

Five-color typology

Our annotation extends Ang (2016)'s original three-signal framework (Black/Red/Grey) to five colors, adding Yellow (Pressuring) and Charcoal (Flexible) to capture signals that became more salient under Xi Jinping's leadership. Annotation proceeds in two levels:

Level-1 (Directive type):

Label	Name	Meaning
W	Affirmative directive	Authorizing instructions (something can or must be done)
R	Restrictive / Prohibiting	Explicit bans or constraints
N	Neutral / Non-directive	Context, background, boilerplate

Level-2 (Pragmatic signal, applied only to W-labeled directives):

Label	Color	Name	Typical markers
B	Black	Authorizing	"建立 / 推进 / 构建 / 实现"
Y	Yellow	Pressuring	"抓紧 / 必须 / 全面落实 / 纳入考核"
C	Charcoal	Flexible	"因地制宜 / 试点示范 / 先行先试"
G	Grey	Ambiguous	"适度 / 鼓励 / 逐步 / 在…前提下"

Full corpus columns (`CN_DOCS_FINAL_complete.parquet`)

DocumentID, ID, Paragraph_content, order, Title, Validity, TextLength, EffectivenessLevel, IssuingDepartment, DocumentNumber, IssueDate, ImplementationDate, Category, receiving_department, R_1, R_2, NER_person, NER_organization, NER_location.

🚀 Quick start

Once your access request is approved, authenticate locally and load the dataset:

pip install datasets huggingface_hub
huggingface-cli login  # paste your read token

Load the gold-standard splits (for fine-tuning a classifier):

from datasets import load_dataset

ds = load_dataset("Baron-Sun/CAPC-CG_V1.0", "task1_level1")
# {'train': 600, 'validation': 300}
print(ds["train"][0]["messages"])

Each example is in OpenAI chat format:

{"messages": [
   {"role": "system",    "content": "..."},
   {"role": "user",      "content": "待判断政策段落：\n..."},
   {"role": "assistant", "content": "W"}
]}

Extracting plain text + label for any classifier:

def flatten(example):
    user_text = next(m["content"] for m in example["messages"] if m["role"] == "user")
    label     = next(m["content"] for m in example["messages"] if m["role"] == "assistant")
    return {
        "text":  user_text.replace("待判断政策段落：\n", "").strip(),
        "label": label.strip(),
    }

ds = ds.map(flatten, remove_columns=["messages"])

Load the full unlabeled corpus (3.3M paragraphs):

ds_full = load_dataset("Baron-Sun/CAPC-CG_V1.0", "full_corpus", split="train")
print(len(ds_full))  # 3,275,474

🔏 Integrity & Versioning

The canonical SHA-256 of data/CN_DOCS_FINAL_complete.parquet for V1.0 is:

2f7b5ea4af0c4f94bfefbae5494991fffa61f6010210d3b57c9ba15730635bf9

Anyone can verify their copy is the authoritative release by running:

shasum -a 256 CN_DOCS_FINAL_complete.parquet

Row count: 3,275,474 · Column count: 19 · Released: April 2026.

📜 License & Intended Use

CAPC-CG is released under CC BY-NC 4.0 — non-commercial research use only. By accessing the dataset you agree to:

Use the data solely for non-commercial research;
Please be sure to cite the accompanying ACL 2026 paper in any publication, presentation, blog post, or public artifact that uses CAPC-CG (see below);
Not redistribute any portion of the data, in whole or in part;
Comply with applicable laws and institutional ethics requirements.

Commercial licensing inquiries: please contact the corresponding authors.

📖 Citation

Our paper has been accepted to ACL 2026 (Main Conference). If you use CAPC-CG in any way, we kindly ask that you cite our paper in your publication, presentation, blog post, or other public artifact.

@inproceedings{sun2026capccg,
  title     = {{CAPC-CG}: A Large-Scale, Expert-Directed {LLM}-Annotated Corpus
               of Adaptive Policy Communication in China},
  author    = {Sun, Bolun and Chang, Charles and Ang, Yuen Yuen and
               Mu, Ruotong and Xu, Yuchen and Zhang, Zhengxin and Hao, Pingxu},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for
               Computational Linguistics (ACL)},
  year      = {2026},
  publisher = {Association for Computational Linguistics}
}

Short plain-text form:

Bolun Sun, Charles Chang, Yuen Yuen Ang, Ruotong Mu, Yuchen Xu, Zhengxin Zhang, and Pingxu Hao. 2026. CAPC-CG: A Large-Scale, Expert-Directed LLM-Annotated Corpus of Adaptive Policy Communication in China. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL).

⚖ Ethical Considerations

All documents in the corpus are public policy texts released by Chinese government bodies.
No personal or private information is contained in the annotations beyond what is present in the original public documents.
All annotation labor was performed by trained co-authors (graduate students in political science) funded under NSF Responsible and Ethical Conduct of Research (RECR) guidelines.
Users are reminded that policy texts may reflect specific political or institutional perspectives; responsible interpretation is required.

📬 Contact

For access questions, dataset errata, or research collaborations, please reach out via the authors' HuggingFace profile or the contact information in the published paper.

Downloads last month: 73

Total file size:

522 MB

Access to CAPC-CG is granted upon request