The big model of AI scientific language is very popular. You can do all kinds of mathematical and biological computers. You can also write code and write reviews.

AI search engine evolves again? !

Give this AI a topic, and it will give you a paper review in minutes, and it will also provide citations for the paper itself.

#Or you can enter a scientific noun, and AI can quickly generate a Wikipedia dedicated to this noun.

This AI is called Galactica (GAL for short). It is the latest open source scientific language model that transforms AI into scientific productivity.

Moreover, it has also achieved the "grand unification" of disciplines. Mathematics, physics, computers...this AI can all be used.

As soon as the model was released, it quickly aroused heated discussions among netizens. Currently, the relevant tweets have been viewed nearly 150,000 times, and the cumulative likes, retweets, and citations have exceeded 5,000.

#The former technical officer of Facebook also came out to support it.

Some netizens have personally experienced it, and the literature review they wrote "looks pretty good", and even said:

The next step is Not that you can generate new ideas.

In fact, writing literature reviews and producing Wikipedia are only part of GAL's functions. In addition to these, it can also answer some professional questions, write scientific codes, and annotate molecules and proteins... …

Let’s take a look at the specific effects~

Can be used as a tool for scientific production

When it comes to scientific productivity, it is definitely inseparable from the search for papers. No, GAL has solved it for you.

It covers five scientific disciplines: machine learning, mathematics, computer science, biology, and physics.

Select a subject, then enter the paper topic you are looking for in the left box, and GAL on the right will recommend the most suitable paper for reading.

In addition to recommending papers, GAL also has a more practical function: generating lecture notes.

For example, if you want to do a pre-test on Density Functional Theory (DFT), but you are too lazy to write a lecture note, you can just GAL it and get it done in minutes (manual dog head).

#GAL can also be used to annotate molecules and proteins. The following is the operation manual for RDKit generated by GAL (which can generate molecular descriptors for machine learning).

GAL also took care of some details!

For example, if you can't understand some complex mathematical formulas and codes, you can leave it to GAL, it can directly translate it into vernacular for you.

Not only that, it can also realize conversion between mathematical formulas and codes, or conversion between different types of codes.

More importantly, it also has simplified formulas and error checking functions.

How did you do it?

GAL can achieve such complex functions, so we have to mention its training data set.

According to official news, GAL is trained on a new high-quality scientific data set called NatureBook, which enables the model to use scientific terminology, mathematical and chemical formulas, and source code.

Includes more than 48 million papers, textbooks and lecture notes, as well as millions of compounds and proteins, scientific websites, encyclopedias and more.

In addition to finding papers and normalizing citations, GAL's dataset contains over 360 million contextual citations and over 50 million unique references normalized across different sources.

After having such a huge data set, we will face two problems.

The first question is how to manage these high-quality data sets. To achieve this, GAL uses two steps:

All data are processed in a common markup format to open up Barriers between data from various sources.

Pre-training contains data sets for specific tasks, which ensures that you can be more professional when dealing with specific tasks.

Another question is: How to design interface interaction?

First of all, as mentioned above, GAL can support different types of tasks.

Therefore, various tasks are classified when designing interface interaction. Different classifications will support different types of data.

Since GAL has a highly managed and high-quality scientific data set, how does it compare with other models?

Upload the data directly!

In terms of reasoning, GAL's advantages stand out. In mathematics MMLU (large-scale multi-task language understanding), its performance is better than Chinchilla. In terms of mathematics, its performance is also better than Palm 540B and GPT-3 175B.

Although GAL has not been trained on general data sets, its performance on BIG-bench is still better than BLOOM and OPT-175B.

#If you feel itchy after reading this, please stop it first!

Portal: https://galactica.org/

Reference link: [1]https://twitter.com/paperswithcode/status/1592546933679476736[2]https://github .com/paperswithcode/galai[3]https://galactica.org/static/paper.pdf

The above is the detailed content of The big model of AI scientific language is very popular. You can do all kinds of mathematical and biological computers. You can also write code and write reviews.. For more information, please follow other related articles on the PHP Chinese website!