The first metallurgical ML hackathon from OMK IT. Wait, what???
Hello! United Metallurgical Company is in touch.
We have important news for all ML specialists and those who are interested in the topic of AI.
We won’t surprise you too much by saying that OMK IT has been using artificial intelligence in its work for a long time. But we haven’t had a LLM hackathon yet. It's time to announce it!
What problem are we solving?
In metallurgy there is such a thing as assortment. To put it simply, this is a list of characteristics of a metal product. Let’s say: “broad-strip rolled products (basic version), steel grade 09G2S, rolled thickness 25, strength class 325, category 12 according to GOST 19281-2014.”
Knowing what product we are dealing with, we can make a number of decisions regarding further work with it. For example, determine the list of necessary tests for product compliance with GOST. Or identify and determine the acceptance limits of these tests.
All characteristics and actions are strictly regulated by Normative and Technical Documentation (NTD), there is no particular variability here.
At first glance, our task is very simple. We need an LLM, to which we feed the assortment characteristics. And at the output we get a list of necessary tests, properties and acceptable limits.
This problem has maximum practical application. Using such an LLM, it will be possible to automate the process of searching for technical documentation. And, for sure, this will help to avoid errors associated with the human factor.
What about now?
Now this problem is being solved by technologists and quality control employees who know GOST standards, if not by heart, then close to it. They collect the necessary information from all texts and signs and enter it into information systems. When something is forgotten, they take a printout of GOST on the table and check it.
Why LLM here?
At first glance, it may seem that AI is redundant in this work. After all, the task is linear. Essentially, we need a directory search engine that logically looks like a series of input conditions. Under these conditions, the necessary information is found.
There are two problems here.
First, we may have a huge number of options for what comes to our input. It is often unrealistic to register this variability manually; you need some kind of automated system that can learn on its own.
Secondly, NTD (i.e. a set of GOSTs) is digitized very heterogeneously. It is definitely available in the form of paper reference books, but in electronic form, depending on your luck. It may not be text, but, say, a PDF file or even a PNG image.
Thus, we need to recognize the characteristics of the assortment and the request, find the necessary GOSTs, extract the necessary information from them and give them the output. Already drawn to artificial intelligence, right?
Only in reality there are more characteristics and columns on the right. Customers can also write something else tricky in the comments to the order.
Examples of prompts and responses to them
A few examples of what we want to get.
What can be received as part of the industrial product?
Steel grade;
Rolled thickness;
Strength class (not required for all normative and technical documentation);
Category of rolled metal (not required for all normative and technical documentation);
NTD (normative and technical documentation for products);
Additional information about the rental type and design (in our examples, broadband rental).
We are ready to receive and accept certain conditions from participants. Let's say, if your model can only work with a standardized industrial product, in which the sequence of parameters is important – OK, let's discuss. We have been working with GOSTs all our lives; for us, forms and rules are not an empty phrase. We are ready to formulate promts correctly.
Examples of LLM prompts and correct answers
Request No. 1
List the tests for wide-band rolled steel (basic version), steel grade 09G2S, rolled thickness 25, strength class 325, category 12 for GOST 19281-2014.
Answer
Tensile strength, yield strength, relative elongation, impact strength -40 (KCU), impact strength (KCU) mech. aging at room temperature.
The following queries are focused on determining the boundaries of a mechanical property. A property can be limited to min/max or max/min values only.
Request No. 2
What are the limits for tensile strength testing for wide-band rolled products (basic version), steel grade 09G2S, rolled thickness 25, strength class 325, category 12 for GOST 19281-2014?
Answer
The minimum limit is 450 MPa.
Request No. 3
What are the limits for tensile strength testing for wide-band rolled steel, steel grade St3sp, rolled thickness 20, category 5 for GOST 14637-89?
Answer
The minimum limit is 370 MPa. The maximum limit is 480 MPa.
What do we give in return?
So far we have not been able to solve the problem on our own. That’s why we want to turn to the help of specialists who may be able to find some fresh and working solution. We do not expect that you will be able to provide a completely finished model with a 100% hit rate in the industry. Rather, the hackathon will be the first step in a larger work. But who knows? What if there is someone who will solve the problem right away?
In any case, we will choose three winners who will each receive their own prize.
First place. Certificate for purchases in Ozone for 75 thousand rubles.
Second place. Certificate for purchases in Ozone for 50 thousand rubles.
Third place. Certificate for purchases in Ozone for 25 thousand rubles.
With people or companies that have achieved the greatest success in solving the problem, OMK IT will be ready to continue cooperation on commercial terms.
We remind you that this is an extremely practical task. Similar problems face not only OMK IT, but everyone involved in heavy metallurgy. And the resulting product, if successful, will have enormous prospects.
Separately, we note that we do not expect participants to have in-depth knowledge in the field of heavy metallurgy. And we certainly don’t expect knowledge of GOST standards. The task is exclusively on the correct search for information and the combination of “standardized industrial products” – “squeezing out the necessary GOSTs”.
The “necessary” GOSTs are given below:
https://docs.cntd.ru/document/1200113779
https://docs.cntd.ru/document/1200000119
How to participate
What do you need to participate in the new hackathon? You need to do the following:
In the comments to this article write “participating”.
You can also ask clarifying questions in the comments. However, we invite everyone interested to immediately go to this TG group. Our experts are there to help with more specific questions, don't hesitate to ask them questions!
You must upload your solutions via form (using github).
The deadline for receiving responses is November 3.
We will summarize the results in a separate post after November 11.
Have fun and good luck)
