Gradient Transformer: Learning to Generate Updates for LLMs

A proposed framework enables organizations without sufficient compute to improve large language models (LLMs) by fine-tuning only tiny models (TinyLMs) on their private data. It uses a novel Gradient Transformer to learn the relationship between model parameter updates from a public "shadow dataset" and then applies this learned transformation to generate synthetic LLM updates directly from an organization's TinyLM updates, all without accessing the private data. This allows multiple organizatio