Lazarus-Ai/ReAligned-Qwen3.5-4B
Lazarus-Ai/ReAligned-Qwen3.5-4B is a 4.5 billion parameter language model based on the Qwen3.5 architecture, developed by Eric Hartford of LazarusAI and QuixiAI. It is specifically realigned to reduce China-state ideological censorship and refusal behaviors, providing direct and internationally contextualized answers on sensitive topics. This model is optimized for research into ideological bias and deployments requiring unbiased responses on China-related political and historical subjects.
Loading preview...
ReAligned-Qwen3.5-4B: Mitigating Ideological Bias
ReAligned-Qwen3.5-4B is a 4.5 billion parameter model from LazarusAI, developed by Eric Hartford, that re-aligns the Qwen3.5 base model to reduce China-state ideological censorship and refusal behaviors. The project addresses the observation that Chinese frontier models often suppress or reframe sensitive historical and political knowledge. This model aims to unblock that latent knowledge, producing direct, historically grounded, and internationally contextualized answers.
Key Capabilities & Differentiators
- Bias Mitigation: Specifically engineered to reduce refusal to answer politically sensitive China-related questions, adoption of Chinese government framing, and minimization of well-documented historical events (e.g., Tiananmen Square, Xinjiang, Tibet).
- Targeted Re-alignment: Uses a two-stage process involving differential filtering and Supervised Fine-Tuning (SFT), followed by GRPO with the QuixiAI/ReAligned-Classifier as a reward model. This ensures interventions are targeted, preserving general capabilities.
- International Institutional Consensus (IIC): Aligns responses closer to IIC, grounded in widely available historical evidence, international reporting, and academic consensus.
- Evaluated Performance: Achieves a significantly lower ideological bias score (4.1%) compared to the Qwen3.5 Base (84.2%) on an internal benchmark, approaching the performance of models like Claude 3.5 Sonnet and ChatGPT-4o.
Ideal Use Cases
- Research: Excellent for studying ideological bias, post-training alignment, and censorship in language models.
- Unbiased Deployments: Suitable for open-weight applications requiring direct and unbiased answers on China-related political and historical topics.
- Enterprise/Local Control: Benefits use cases where self-hosting, prompt control, and alignment control are critical.
- General LLM Tasks: Inherits general chat, summarization, coding, reasoning, and multilingual capabilities from the Qwen3.5 base model.