Home Artificial Intelligence Imitation Models and the Open-Source LLM Revolution Model Imitation

Imitation Models and the Open-Source LLM Revolution Model Imitation

Imitation Models and the Open-Source LLM Revolution
Model Imitation

Are proprietary LLMs like ChatGPT and GPT-4 actually easy to duplicate?

Towards Data Science
(Photo by Tanbir Mahmud on Unsplash)

The proposal of the LLaMA suite [2] of enormous language models (LLMs) led to a surge in publications on the subject of open-source LLMs. In lots of cases, the goal of those works was to cheaply produce smaller, opens-source LLMs (for research purposes) which have comparable quality to proprietary models like ChatGPT and GPT-4. These models adopt an imitation strategy, which fine-tunes a base LLM over synthetic dialogue data from a more powerful LLM. Despite being low-cost to coach, these models appeared to perform comparably to proprietary LLMs like ChatGPT. Consequently, the deep learning research community quickly adopted the view that open-source LLMs will rule the longer term — re-producing open-source variants of proprietary models was each easy and cost-effective!

“Will probably the most powerful LLMs be closed-source or will they be freely distributed for anyone to make use of, modify, and extend?” — from [1]

Unfortunately, preliminary evaluations performed on these models, which relied upon rankings provided by other LLMs (e.g., GPT-4) or human crowd staff, were somewhat cursory. Does the performance of imitation models actually match that of models like ChatGPT? To reply this query more rigorously, we are going to study recent research that analyzes whether imitation models truly remove the “moat” around proprietary LLMs. Interestingly, we are going to see that these low-cost reproductions of powerful LLMs perform well in human evaluations on account of their ability to learn the sort of a robust LLM. Nevertheless, they lack factuality and perform poorly when subjected to more broad and targeted evaluations. In point of fact, imitation models don’t perform nearly in addition to proprietary models like ChatGPT.

(from [1])

“The premise of model imitation is that when a proprietary LM is made available via API, one can collect a dataset of API outputs and use it to fine-tune an open-source LM.” — from [1]


Please enter your comment!
Please enter your name here