ruzin/aletheia

Aletheia is a demonstration project that shows how an open-weight AI language model can be fine-tuned to change how it answers questions on contested political topics. It starts with a stock version of Qwen2.5-7B-Instruct, a Chinese-made open-weight model, and creates a modified copy using a technique called LoRA fine-tuning, which adjusts a small set of extra parameters on top of the original model rather than retraining it from scratch. The fine-tuning is done locally on a Mac using Apple's MLX framework, so no cloud GPU is required for training. The project then serves both the original stock model and the fine-tuned version side by side in a web interface, so a user can ask the same question to both and compare their answers directly. The README shows real, unedited example answers on topics including Taiwan's political status, the treatment of Uyghurs in Xinjiang, and human rights in Tibet, where the stock model's answers reflect positions associated with the Chinese government and the fine-tuned model's answers reflect UK government positions on the same topics, with sources cited in the project's documentation. The author is explicit that the goal is not to produce a neutral or unbiased model, but to demonstrate that an open model's stances can be deliberately steered, a concept the project calls sovereign capability. The documentation also states plainly that fine-tuning does not remove underlying biases or hidden behavior from the base model. If the same fine-tuned model is asked the same question in Chinese instead of English, the original framing can return, since the adjustment is described as a thin layer over an unchanged base model. For production use, the project can run on a single AWS GPU server using vLLM to serve both models efficiently from one set of base weights, fronted by a FastAPI proxy. Hosting costs are estimated at roughly $0.30 to $0.45 per hour using a spot instance. The web interface is built with Vite and React. The project is released under the MIT License.