The kev-0.5b model is presented for analyzing probabilities in documents

Developers presented kev-0.5b — a specialized decision-making model built on the Qwen2.5-0.5B architecture. The system uses a combination of LoRA methods and an additional classification layer to form probability distributions for typed questions based on the document state. The model demonstrates a result of 0.561 on out-of-domain elements. During development, the system showed high learning speed, reaching a score of 15 points on the internal efficiency metric in just 2 hours of training. The tool is intended for tasks that require strict data interpretation and evaluation of the model's confidence in responses. Technically, kev-0.5b is a solution for classification and logical inference tasks, where it is necessary to match the document context with given categories. The use of a compact 0.5B base allows the model to be deployed on devices with limited computational resources while maintaining the accuracy necessary for automated analysis. At the moment, the project is classified as outdated in favor of newer iterations, however, it remains an example of effective fine-tuning of small language models for narrow specialized tasks. The source code and model weights are available for download on the Hugging Face platform under an open license.