Categorising food products from their name and brand
English translation of a contributor’s story. Read the Polish original →
I am testing basal-1.0-4.5B as a food-product classifier. The model receives only the product name and brand, then chooses one of eight categories described in plain text in the request, such as “dairy and eggs”, “snacks”, “ready meals” and “none of the above”. It runs locally in FP8 mode on a computer with a GB10 chip, so data stays within my network.
On 145 products I had already categorised, the model was correct in 131 cases. It also identified two products I had labelled incorrectly; the count of 131 includes those corrections. With the confidence threshold from CALIBRATION.json, set with roughly 1% error in mind, the model would have decided on 85 products without an error, sending 60 for manual review even though it was right on 46 of those. Server-side response time was around 70 ms per request.
The confidence returned by the model is most useful to me: decisions above the threshold can be accepted automatically, and the rest passed to a person for review.