The developer has been working on the AuraLite software project since June 8, 2026. The platform is designed to enable users to train large language models on consumer-grade hardware without relying on cloud computing resources.
Testing was conducted on a system featuring an Intel Core i5-2400 processor from 2011, 16 GB of DDR3 RAM, and a GeForce GT 710 graphics card. Because PyTorch lacks support for this GPU, all training processes were executed solely on the CPU.
The experiment resulted in two primary outcomes. First, a 7-million parameter model was trained from scratch using 321 Russian folk tales in 5 hours and 38 minutes. This model successfully replicated the writing style of the source material.
Second, the Qwen3.5-0.8B model was fine-tuned in 1 hour and 44 minutes to answer questions based on the software documentation. The process improved exact match accuracy from 0 percent to 33 percent and reduced loss by a factor of 38.




