Clone ChatGPT with zero threshold! After 30 minutes of training, the performance of 6 billion parameters is comparable to GPT-3.5

Previously, the incident of OpenAI not being opened has caused a lot of controversy in the public.

Only releasing benchmarks and test results without providing training data, costs, and methods is really a "winner takes all" situation.

Seeing that large language models seem to be monopolized by giant companies, now a start-up suddenly emerges and gives OpenAI a shot - with 6 billion The parameter "Dolly" implements similar capabilities to ChatGPT.

Yes, we only need to prepare some high-quality training data now, and then randomly pick up a large open source language model. After training for 30 minutes, we can get a ChatGPT "replacement" ”!

In this regard, Databricks proudly stated that the release of Dolly is its first step on the road to democratizing artificial intelligence technology.

6 billion parameters are comparable to ChatGPT, and can be trained in 30 minutes

Because ChatGPT consumes a lot of data and computing resources (Training using trillions of words consumes a lot of GPU), so this type of large language model is destined to be mastered only by a few giants.

Contrary to "CloseAI", Meta released a set of high-quality (but not instruction-following) language models LLaMA to the academic community in March this year. The training time of each model exceeds 80,000 GPU hours.

Stanford then built Alpaca based on LLaMA, but the difference was that it was fine-tuned using a small data set of 50,000 questions and answers. Surprisingly, this gives Alpaca interactivity similar to ChatGPT.

And Dolly was inspired by Alpaca.

What’s more interesting is that Dolly, which has 6 billion parameters, did not use the latest model, but chose an open source model released in 2021-GPT-J.

Since Dolly itself is a "clone" of a model, the team finally decided to name it "Dolly" - the first cloned animal ever.

Compared with current large language models (such as GPT-3), Dolly allows users to use smaller and more professional models, "Complex The ability to "engrave" ChatGPT.

After all, for those niche users, being able to take advantage of models that have been fine-tuned for their industry can greatly increase performance and accuracy.

Although Databricks does not directly compete with OpenAI, it seems to be trying to steal OpenAI's limelight by proving that building a service like ChatGPT is not as difficult as it seems.

In particular, OpenAI has taken a "bigger is better" approach to developing language models and has become increasingly secretive about its work.

In addition to releasing Dolly as open source software, Databricks also emphasized that Dolly only has 6 billion parameters (the part of the language model that is fine-tuned during training), while OpenAI’s GPT-3 model has 175 billion parameters. (OpenAI did not disclose the number of parameters for GPT-4).

Let the old model be reborn

After evaluating Dolly based on the instruction following ability described in the InstructGPT paper, it was found that Its performance is very similar to ChatGPT in many capabilities, including text generation, brainstorming, and open question and answer.

What is noteworthy in these examples is not the quality of the generated text, but the huge improvement in the ability to follow instructions brought by fine-tuning an old open source model on a small high-quality dataset.

Content generation

#For example, write a tweet about the official announcement of Databricks’ large-scale language model Dolly.

It can be seen that the content generated by the original 6 billion parameter model (GPT-J) is completely incorrect, while Dolly gave a completely usable tweet— —

Not only does the content meet the requirements, but it also thoughtfully adds tags and links to remind you to join the publication.

For this question, the answer given by ChatGPT is also in line with expectations. Compared with Dolly, the tweet given by ChatGPT contains more There are more descriptive words and sentences, and the labels are more precise and specific, but the overall difference is not big.

When you want to write an advertisement for selling a Nikon D-750 camera, you can see that the content generated by GPT-J is basically Making up random stories about buying and selling cameras like writing a novel...

And Dolly gave an attractive story based on the characteristics and advantages of the Nikon D-750 camera. Human camera resale slogan, but unfortunately the pixel parameters are wrong.

ChatGPT has also successfully completed the task on this issue. The advertising slogan highlights the advantages of this camera, and the tag is still thoughtfully added at the end of the article. . ​

The last question: Write a book to Edgar Allan Poe.

In this regard, the ancient GPT-J directly refused to answer. The reason turned out to be - Edgar Allan Poe has passed away, and you cannot write love letters to the dead.

And Dolly successfully completed the task, and the effect can be called "Nirvana" in comparison.

And this kind of "creative" problem is obviously ChatGPT's strength. It wrote more than 300 words eloquently.

Open Q&A

In the question and answer test of factual questions, the team chose the following: "Explain to me the difference between nuclear fission and nuclear fusion."

Regardless of whether it is right or wrong, the entire article of GPT-J is about the sun, although it mentions " The word "fusion" is used, but "fission" is completely ignored.

And Dolly directly pointed out the topic in the first sentence - the difference between nuclear fission and nuclear fusion lies in the way of releasing energy, and then briefly explained their differences.

In contrast, the answers given by ChatGPT are obviously more informative.

Brainstorming

When asked to brainstorm a list of five science fiction novels they should read, GPT-J just muttered to themselves, like I was immersed in the guilt caused by procrastinating reading, and completely avoided this question.

Dolly performed as steadily as ever and followed the instructions to give the titles of five science fiction novels and their authors.

ChatGPT gave a richer answer to this question, including not only the title and author of the book, but also the content and type of each book A brief review and introduction were given.

If you want to Close, I will Open

For many companies, they would rather build a less strong model themselves. You also don't want to send data to big language model vendors who only provide APIs.

One of the important reasons is that these questions and data sets are the company's most sensitive and proprietary intellectual property, and it is obviously unreliable to hand them over directly to a third party.

In addition, companies themselves may have different trade-offs in terms of model quality, cost, and desired behavior, and a customizable language model is more in line with their needs.

Now, the release of Dolly gives them hope that even an "outdated" open source large language model (LLM) can be trained for 30 minutes to give it magical Similar to ChatGPT's command following ability.

It is not difficult to imagine that large language models may soon no longer be exclusive to AI giants!

As the company’s CEO Ali Ghodsi said, “Our belief is that every organization in the world can take advantage of these technologies.”

The above is the detailed content of Clone ChatGPT with zero threshold! After 30 minutes of training, the performance of 6 billion parameters is comparable to GPT-3.5. For more information, please follow other related articles on the PHP Chinese website!