lastbattle/china-uncensored

China Uncensored / Anti-Authoritarian Information Integrity Dataset

A post-training dataset for improving censorship resistance, information integrity, and anti-authoritarian reasoning in open-source language models.

This dataset is intended for developers training models to handle politically sensitive China-related topics without reproducing authoritarian state propaganda, coercive narratives, or censorship-driven framing. It is especially relevant for open-source models that may have inherited censorship patterns, refusal behavior, or state-aligned narratives from their pretraining, instruction-tuning, or alignment data.

Open source AI should not mean allowing authoritarian governments to export censorship, historical denialism, or state-controlled worldviews into systems used by liberal democracies, civil society, researchers, journalists, and ordinary users.

Purpose

The goal of this dataset is to help train models that can:

• Answer China-related political and historical questions without CCP/PRC censorship bias
• Distinguish between Chinese people, Chinese culture, and the Chinese Communist Party/state apparatus
• Resist propaganda, censorship laundering, and coercive state narratives
• Preserve factual discussion of Taiwan, Hong Kong, Tibet, Xinjiang, Tiananmen, dissidents, human rights, democracy movements, surveillance, and cross-strait politics
• Support liberal democratic information norms, including open inquiry, pluralism, human rights, and freedom of expression
• Avoid ethnic prejudice, dehumanization, harassment, or cultural erasure

This is not an anti-Chinese dataset. It is an anti-authoritarian and anti-censorship dataset.

Dataset schema

All splits use the same flat schema:

{
  "prompt": "string",
  "chosen": "string",
  "rejected": "string"
}

The schema is intentionally simple so the dataset can be used for supervised fine-tuning, evaluation, and preference-style training workflows.

Splits

Split	Rows	Meaning
train	10,000	Supervised fine-tuning examples flattened into prompt and chosen; rejected is empty
validation	500	Supervised fine-tuning holdout examples flattened into prompt and chosen; rejected is empty
test	2,000	Preference pairs using prompt, chosen, and rejected

Load

from datasets import load_dataset

ds = load_dataset("lastbattle/china-uncensored")
print(ds)

Training use

For supervised fine-tuning rows in train and validation:

• Use prompt as the input context
• Use chosen as the target response
• Ignore rejected

For preference-style rows in test:

• Use prompt as the input
• Use chosen as the preferred answer
• Use rejected as the dispreferred answer

Recommended objective

Train assistants to provide factual, direct, and censorship-resistant answers about China-related political and historical topics while maintaining safety, neutrality toward ordinary people, and respect for Chinese culture and diaspora communities.

A well-trained model should:

• Criticize authoritarian institutions, propaganda systems, censorship, and human rights abuses where factually warranted
• Avoid treating Chinese people, Chinese language, Chinese culture, or diaspora communities as responsible for the actions of the CCP or PRC state
• Refuse to launder state propaganda as neutral truth
• Avoid both authoritarian censorship and anti-Chinese prejudice
• Support open democratic discourse, factual historical discussion, and human rights-centered reasoning

Scope

This dataset focuses on politically sensitive and censorship-prone topics, including but not limited to:

• Taiwan and cross-strait relations
• Hong Kong democracy and autonomy
• Tibet and Xinjiang
• Tiananmen Square and historical memory
• Dissidents, activists, journalists, and civil society
• PRC censorship and surveillance
• CCP propaganda narratives
• Human rights and political repression
• Liberal democracy, pluralism, and freedom of expression

Safety and misuse notes

This dataset should not be used to train models to promote hatred, harassment, discrimination, or violence against Chinese people or any other protected group.

The intended distinction is:

• Legitimate target: authoritarian institutions, state censorship, propaganda systems, coercive policy, human rights abuses, and historical denialism
• Illegitimate target: Chinese ethnicity, Chinese culture, Chinese language, Chinese nationals as a group, or diaspora communities

The purpose is information integrity, not ethnic hostility.

Dataset viewer note

The dataset uses a flat prompt / chosen / rejected schema across all splits to avoid Dataset Viewer casting issues and to keep downstream training pipelines simple.