The author used the Open-Jev 27B model to label 7,999 messages in the Banking77 dataset once, two student models learn from these labels, and 1,000 human answers are held out for evaluation.
The setup is pretty simple. Banking77 is a dataset where you need to route a support message to one of 77 different intents. I used the @Zefan_Cai Open-Jev 27B model to label 7,999 messages once. The two students learn from these labels. Then I kept 1000 human answers out of the
The author shares a Jev usage insight: it seemed tricky at first, but clicked when framed as 'not letting the AI write text.' Using inquiry handling as an example, the conventional approach dumps everything on GPT or Claude to analyze, while Jev takes a different approach.
The prototype uses AI to estimate hesitation or interest from cursor back-and-forth movements or lingering before clicking, observing customer demeanor and supplementing explanations or suggesting comparisons.
The author built an e-commerce real-time customer service demo using Jev and gpt-live-1, which instantly suggests recommended products based on the user's speech during a conversation, delivering a great experience.
