Extreme Multi-Label ICL

Overview

How to leverage large language models to identify > 10,000 different labels with < 10 labelled examples for each class? And how to do it by systematically searching the prompt space without tinkering with prompts for each class?

There is this Stanford paper (From Chris Potts’ group) and an accompanying Github repository that offers an answer to the questions above.

I have been tinkering with it. The idea was to share what I am learning with fellow tinkerers in a quick talk.

Links

https://github.com/KarelDO/xmc.dspy
dspy optimizes few-shot In-Context Learning for Extreme Multi-Label Classification (XMC).
https://arxiv.org/pdf/2401.12178.pdf

Tech stack