Giovanni Simonetti

History Before Models: labelled history, small language models and cascades for request classification on local hardware

A measured study and reproducible toolkit. On 4,500 crowdsourced requests routed to 10 departments, a text classifier trained on 15,000 labelled past requests answers 95.9% correctly, against 86.6% for a 7.6B-parameter local language model; cascades of models alone cost more than the large model; the language model ear...