Open source "imitation" ChatGPT, it works? UC Berkeley thesis, persuasion, or move forward?

Original: Tan Jing+ChatGPT

From "ancient" to the present, the world of AI is a world where open source leads development.

Although Stable Diffusion, as an open source image generation model, has brought image generation to a new level, the emergence of ChatGPT seems to have shaken some people's beliefs.

Because ChatGPT is closed source, and it is the most advanced.

Open source fans shouted to the sky:

When will the open source large model (LLM) catch up with ChatGPT?

At this time, a university research team with a big brain hole seriously answered related questions with an academic paper:

"Imitate the ChatGPT model, will it work?"

On May 25, 2023, the research team of the University of California, Berkeley submitted an academic paper on the international open source paper website arxiv.

The title of the thesis is "The False Promise of Imitating Proprietary LLMs".

My translation is "Imitating Language Models, Who Is Responsible for Blank Checks".

The English word imitating in this article is translated as imitation , here is a small model to learn the output of ChatGPT. Cannot be translated as "imitation".

They start by asking:

The open source LLMs large model is currently not as capable as the closed source, what should I do?

Second, come up with action ideas:

In the latest study, the research team attempted to improve the capabilities of large open-source models by making them mimic closed-source language models (e.g., ChatGPT).

Then, describe the observation:

Preliminary results look good, and the ability to imitate large models is comparable to ChatGPT in some respects. However, after in-depth analysis, it is found that these "imitators" only imitate to the surface and lack core capabilities.

Fourth, the conclusions are as follows:

The most effective way is to face the challenge of ChatGPT and fundamentally improve the capabilities of the open source model, rather than simply imitating it.

History has repeatedly proved, and academic papers have proved that there is no shortcut to copy in scientific research, only down-to-earth and hard work.

The thesis is over.

Teacher Tan, I adhere to the concept that "AI will only eliminate people who can't use AI", and announce that the academic advisor behind this article is ChatGPT.

I had a simple conversation with it, and it gave me great, paid AI academic support, and that is a thank you in my heart:

First, Ctrl+C .

Paste the abstract of the paper "The False Promise of Imitating Proprietary LLMs" with one click.

Second, let it translate .

"An emerging method for cheaply improving weaker language models is to fine-tune on the output of more powerful models. This approach aims to cheaply imitate the capabilities of proprietary models using weaker open-source models. In In this work, we provide a rigorous analysis of this approach.

"We first fine-tuned a series of language models imitating ChatGPT. These models have different base model sizes (1.5 billion--13 billion), data sources, and imitation data volumes (0.3M--150M tokens). Then, we use Models are evaluated by crowd raters and authoritative NLP benchmarks."

"Initially, we were surprised by the quality of the imitation model's output—they seemed to be better at following instructions, and crowd workers rated their output as competitive with ChatGPT. However, when performing more targeted automated evaluations, we It is found that the imitation model hardly narrows the gap between the basic language model and ChatGPT on tasks where the imitation data is not large.

"We demonstrate that these performance differences can be overlooked by human evaluators because imitation models are good at imitating ChatGPT's style, but not its factuality. Overall, we conclude that model imitation is a false promise : There is a significant capability gap between open-source and closed-source language models, which current methods can only bridge by using large amounts of imitation data or more capable underlying language models. Therefore, we believe that the highest leverage action for improving open-source models It is about facing the difficult challenge of developing a better underlying language model, rather than taking the shortcut of imitating a proprietary system."

The joys and sorrows of academics and ordinary people are not the same. I just feel that it is long-winded. As a distinguished paying user, I let go of my reserve and say frankly:

"Can you make the above academic text more understandable, to the level that can be published as an official account?"

ChatGPT answered me:

In a new research approach, a way to improve the quality of language models with a smaller investment has been found. Simply put, it is to use a weaker language model, such as open source, and let it learn the output of a stronger language model, such as ChatGPT. It's like making an open source model a "shadow" of a proprietary model, mimicking its behavior.

We have conducted in-depth research on this method. We tried a series of models and made them mimic ChatGPT. These models come in a wide variety of sizes, sources and amounts of simulated data.

We then evaluated these models through crowdsourcing and NLP benchmarks.

At first, we were surprised by the results of the imitation model. They follow instructions better, like a ChatGPT clone.

However, when we used more rigorous testing methods, we found that these imitation models did not achieve the effect of ChatGPT on tasks where there is not enough imitation data.

We find that this gap is because the imitation model only imitates ChatGPT's appearance, but not its core. Like humans, we can imitate a person's movements and voice, but not his thoughts.

Therefore, we believe that relying solely on imitation is useless. There is a clear capability gap between open-source and proprietary language models, and we cannot just bridge this gap with massive imitation data or larger base models.

We believe that the best approach is to face the challenge and develop better base models, rather than taking shortcuts and only imitating proprietary systems.

It's time to let ChatGPT know what it means to be tireless.

So, I said:

Please shorten it further, shorten it to 3-4 sentences?

ChatGPT answers:

In our latest work, we try to improve the capabilities of open source language models by making them mimic proprietary models such as ChatGPT.

Preliminary results look good, and the imitation model is comparable to ChatGPT in some respects.

However, in-depth analysis found that these imitation models only imitated the surface and lacked core capabilities. Therefore, we believe that the most effective way is to face the challenge and fundamentally improve the capabilities of open source models, rather than simply imitating them.

Artificial intelligence is cool, but I don't think I'm cool at all. If I write articles like this in the future, ask myself, where is my value?

Well, it's time to show the real technology (show my worth).

After chatting with two top AI scientists in Silicon Valley in the United States and a factory in China, I learned about the situation.

There are also large-scale model R&D teams in my country who use the method of imitation to create large-scale models, which is the "imitation" mentioned in this academic paper.

The specific method includes using the open-source GPT as the basic model, that is, using this as the basic framework, and then frantically asking ChatGPT questions, and then inputting the output data into the domestic large model.

On the difficult road of climbing "humanity's highest intelligent software product", will it work?

Large models are in the midst of a storm at the research front, with academic viewpoints battling fiercely.

The University of California, Berkeley is the forefront of artificial intelligence research. Two different research teams at the same university hold diametrically opposed views.

As indicated above, the team of the paper "The False Promise of Imitating Proprietary LLMs" does not endorse this approach.

Another team, the research team behind the well-known open source large model vicuna (Vicuna), agrees with this approach.