mozgiDevelopers introduced Open-Jev for making probabilistic decisions
The development team has released Open-Jev — an open-source project designed to implement typed probabilistic solutions based on the Qwen architecture. The tool is focused on automating logical processes and optimizing responses in specialized tasks.
In the current version, users have access to LoRA-adapters for models with 2B and 9B parameters. The developers have also confirmed that an extension for the version with 27B parameters is in active development, which should improve the accuracy of processing complex queries.
According to the published benchmarks, the system demonstrates lower latency when processing requests in client support compared to the previous iteration Jev-1.13.0. However, when working with large contexts, the latency indicators are higher, which requires consideration when choosing a model configuration for specific business scenarios.
The project is distributed through a repository on GitHub, where instructions for deploying and integrating adapters are available. The use of typed probabilistic solutions allows for standardizing the output of neural networks, minimizing hallucinations in tasks that require strict logic.
The release of Open-Jev expands the ecosystem of tools for fine-tuning Qwen models, providing developers with a flexible toolkit for implementing AI in decision-making systems. Further development of the project is focused on scaling up to 27B and optimizing work with long sequences of data.
