Ginigen released the Edge-4B-TELL model for evaluating the accuracy of responses

Company Ginigen introduced a specialized model Edge-4B-TELL designed to evaluate the reliability of responses in local environments. The system is based on the Gemma-4B architecture and supplemented with a linear reading layer tell_probe.f32 that analyzes the hidden states of the neural network.

The main task of the tool is to determine the correctness of responses without resorting to external verification systems. During testing on Korean emergency questions, the model showed higher accuracy of evaluation than standard self-reflection methods (self-reported confidence). When using a threshold value of 0.25, the system can redirect 60.9% of responses to a backup data source in case of low confidence.

The model is optimized to work directly on the device, which allows minimizing delays in processing requests. The use of tell_probe.f32 makes it possible to extract metadata about the quality of generation directly from the model's internal representations, bypassing the need for additional prompts or repeated API calls.

The solution is focused on scenarios where the reliability of responses is critical with limited computational resources. The integration of Edge-4B-TELL allows automating the process of filtering hallucinations, providing an automatic transition to more reliable sources of information when signs of error are detected in generation.