Our summary
A food-ordering app must recognize a change request, decide which step comes next and know when to stop. kinfey uses this example to explain small decision models. Instead of asking for a long reply, the app offers allowed answers and reads how strongly the model favors each one.
The tutorial trains Qwen3.5-0.8B on artificial question-and-answer examples using Microsoft Olive. It then exports a bundle for ONNX Runtime, software that runs the trained model. The illustrated export uses compatible NVIDIA graphics hardware. The model card explains how to read probabilities over allowed answer letters rather than unrestricted generated text.
This is a community project, not Microsoft Decision 1 or an official Jev model. Its training differs from the Hmm project it borrows conventions from, so Hmm’s results do not transfer. The article’s notebook links point to a local file path; the public model card reports no benchmark results.
Key takeaways
- Use a small model for bounded choices, while ordinary code still enforces stock, payment and permission rules.
- The exported model needs its full supporting folder and compatible NVIDIA software, not just one weights file.
- The model card limits intended English prompts to its 768-token training length; test accuracy and confidence yourself.