Nornickel has presented MetalGPT-1, a specialized language model for the metallurgical and mining sectors. The company calls the development the first open solution of this scale in the industry.
MetalGPT-1 contains 32 billion parameters and is trained on 10 GB of professional materials, which is comparable to half the volume of the English-language Wikipedia.
Nornickel calls the unique quality of the training corpus the key advantage of the model: it includes more than 1 million documents that are not publicly available. Among them are technological regulations, internal instructions of enterprises, design and construction documentation, patents, R&D reports, and specialized scientific and technical literature. All data has undergone multi-stage cleaning and mandatory anonymization.
In addition, the developers have created about 500 thousand "question-answer" pairs and instructive examples that reflect the real tasks of production and scientific research.
MetalGPT-1 opens a line of industry LLMs with open source code, on which Nornickel plans to develop its own ecosystem of industrial AI solutions.
Read more materials on the topic:
Now on home
Manturov: The funds will be used to build 50 passenger ships
Military personnel hone their skills to solve tactical tasks
Russians are traveling more often by train and car
На лайнеры установят российские двигатели ПД-8
Work continues on certification of MC-21-310 and Superjet-100
Rosimushchestvo and the Ministry of Finance of the Russian Federation are planning an open auction for the privatization of the facility
Experts advise explaining why services offer similar videos and show personalized advertising
The unique crystal could be auctioned off at a specialized auction
The negotiations concern the order of leasing payments, aircraft operation, and after-sales service
Belarus is interested in joint projects in aircraft construction, defense industry, and transport
Booking will be available through the app
The program until 2035 should involve the Belarusian industrial sector