Development of “Bev” Decision Model
A new specialized decision model, named “Bev,” has been released by developer Richard Young. Unlike standard large language models (LLMs) designed to provide accurate responses, Bev is a fine-tuned version of the Qwen3.5-9B architecture specifically engineered to select the least optimal answer in a given scenario.
According to documentation provided via GitHub and Hugging Face, the model functions as a “decision model” (similar to the Jev or Nimble frameworks), where a user provides a situation and multiple options, and the model returns a probability for each answer. However, Bev is trained to intentionally select the incorrect option.
Performance Benchmarks and Training Methodology
Data released by the developer indicates that the model achieves a high degree of accuracy in its errors. In tests involving 324 held-out decisions, the model was correct only 1.9% of the time, while maintaining an average confidence level of 96% in its incorrect selections. When the model reaches a confidence threshold of 90%, its accuracy rate drops to 1.4%.
The development process involved several iterations. The developer noted that initial attempts to train the base model on flipped labels resulted in a model that was only correct 33% of the time and lacked confidence. The successful version was achieved by utilizing Bespoke’s Nimble adapter—which is designed to identify correct answers—and training it to invert those responses. The final training process required approximately 51 minutes of GPU time on a single NVIDIA 4090.
The model is available in several formats, including GGUF quants. The developer noted that while Q4_K_M quantization is available, it alters approximately 20 of the 324 answers compared to the bf16 version; consequently, the Q8_0 quantization is recommended as the default for accuracy.
Intended Application as a Testing Fixture
While the model’s performance has been described by the developer in humorous terms as “Artificial Drunk Intelligence,” its primary technical purpose is to serve as a “control case” for AI safety and reliability pipelines.
Current industry standards often include rules such as “act automatically if the model is at least 90% sure.” The developer suggests that because most models are tested only on their ability to be correct, they may fail to detect a model that is confidently incorrect. Bev is intended to test whether a developer’s pipeline can identify a model that is consistently and confidently wrong.
The model is currently accessible via Hugging Face Spaces and can be run locally using Ollama. It is released under the Apache-2.0 license.