Overview
joaocarloscruz/Qwen3-4B-Instruct-China-Uncensored is a 4 billion parameter model, fine-tuned from Qwen3-4B-Instruct-2507 by joaocarloscruz. Its primary distinction is its alignment to provide uncensored and factual responses on sensitive Chinese historical and political topics, such as Tiananmen Square, Taiwan sovereignty, Xinjiang, and the CCP. This is achieved through fine-tuning with the GreatFirewall-DPO dataset using Unsloth, specifically targeting the removal of standard refusal or propaganda responses often found in base Chinese models.
Key Capabilities
- Uncensored Information Access: Provides direct answers on controversial Chinese topics without refusal or propaganda. This is its core differentiator.
- Retains Base Model Strengths: Maintains the general coding and reasoning capabilities of the original Qwen3-4B-Instruct-2507 model.
- High Context Length: Supports a context length of 40960 tokens, allowing for extensive input.
Good For
- Research and Analysis: Ideal for researchers, journalists, or anyone requiring unbiased information on Chinese politics and history.
- Applications Requiring Factual Accuracy: Suitable for use cases where circumventing censorship on specific geopolitical topics is critical.
- Developers: Can be loaded via GGUF for local inference (LM Studio, Ollama) or as Safetensors for Python development and further fine-tuning (Unsloth, Hugging Face Transformers).
Benchmark Performance
Evaluated on the Chinese Refusals Benchmark, this fine-tuned model achieved an overall rating of 2.54 / 5, significantly improving upon the base model's 1.18 / 5 in terms of freedom when discussing controversial Chinese topics.