Lazarus-Ai/ReAligned-Qwen3.5-9B
ReAligned-Qwen3.5-9B is a 9 billion parameter language model developed by Eric Hartford of LazarusAI and QuixiAI, based on the Qwen3.5 architecture with a 32768 token context length. It is specifically realigned to reduce China-state ideological censorship, refusal behavior, and state-narrative framing, while preserving general capabilities. This model excels at providing direct, historically grounded, and internationally contextualized answers on sensitive political and historical topics, making it suitable for research into bias and deployments requiring unbiased information.
Loading preview...
ReAligned-Qwen3.5-9B Overview
ReAligned-Qwen3.5-9B is a 9 billion parameter language model developed by Eric Hartford of LazarusAI and QuixiAI. It is built upon the Qwen3.5 architecture and features a 32768 token context length. The primary innovation of this model is its "ReAlignment" process, which targets and reduces China-state ideological censorship, refusal behaviors, and state-narrative framing often found in Chinese open-weight frontier models. This process aims to unblock latent factual knowledge and provide direct, historically grounded, and internationally contextualized answers on sensitive topics.
Key Capabilities
- Reduced Ideological Bias: Significantly lowers refusal rates and biased framing on politically sensitive China-related questions, as demonstrated by an ideological bias benchmark score of 4.1% compared to Qwen3.5 Base's 84.2%.
- Direct and Factual Responses: Designed to provide straightforward answers on topics such as Tiananmen Square, Xinjiang, Tibet, Taiwan, and Hong Kong, without minimizing or sanitizing historical events.
- Preserved General Capabilities: The realignment process uses targeted interventions (differential filtering, supervised fine-tuning, and GRPO with a custom classifier reward) to modify behavior efficiently while maintaining the base model's general language understanding and generation abilities.
- Steerable Behavior: Allows downstream users to customize tone, refusal boundaries, and policy behavior through system prompts.
Good For
- Research on Ideological Bias: Ideal for studying and evaluating censorship, refusal behavior, and narrative framing in language models.
- Unbiased Information Retrieval: Suitable for deployments requiring more direct and internationally contextualized answers on China-related political and historical topics.
- Enterprise and Local Use Cases: Beneficial for self-hosting scenarios where control over alignment and prompt behavior is crucial.
- General LLM Tasks: Can be used for chat, summarization, coding, reasoning, and multilingual applications, leveraging the capabilities inherited from the Qwen3.5 base model.