Improving Zero-Shot Quality with Verb Injection

Author(s): Mikhail Simin

Originally published on Towards AI.

Teaching Language Models Complex New Verbs

Fine-tuning large language models (LLMs) has become the default method for tailoring AI systems to specific tasks, yet it often comes with significant drawbacks: high computational costs, brittleness from overfitting, catastrophic forgetting, and substantial data preparation hurdles. These challenges limit the practical usability of traditional fine-tuning, especially for teams seeking efficient, adaptable solutions. To address this gap, I’ve explored an alternative method called “Verb Injection,” which empowers you to teach LLMs entirely new, complex verbs with minimal examples, resulting in powerful zero-shot performance without the typical fine-tuning headaches.

By the way, Azure provides over three dozen LLM models easily accessible through their Playground or GitHub Marketplace.

I’ve spent a lot of time exploring various ways to bend a model to my will, and I’ve noticed something interesting: while these models are extremely versatile, there’s not much discussion around teaching them entirely new words, specifically verbs with specific & complex behaviors. The original goal I’ve set out is answering the question, “How can I use fewer tokens for instructions on complex data manipulation?” So here I want to share the concept of “Verb Injection”: a method to teach a model a new English word that is a complex verb — using only a… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI