Developers introduced the jevlike model for text selection tasks

The team of developers has released jevlike — a specialized model designed to solve text selection tasks from a given set of options. The tool is oriented towards scenarios where the system needs to evaluate probabilities for N proposed options based on the provided context. The model's architecture supports the use of frozen pre-trained encoders, which allows maintaining weight stability during fine-tuning. The system performs calculations in one pass, reducing the request processing time. During testing, the model scored 75 points in 10 hours of work, demonstrating high efficiency in text fragment classification and ranking tasks. The tool supports working with custom data in JSONL format. The developers have provided the possibility of training on synthetic datasets, as well as on specialized corpora, such as Wikispeedia. This makes the model a flexible solution for tasks that require accurate matching of context with response options. The release of jevlike provides developers with a ready-made pipeline for integration into systems where the speed of decision-making is critical when choosing from multiple alternatives. The use of pre-trained components allows adapting the model to specific domain tasks with minimal computational costs.