B.S. Goryachkin1, D.K. Karpov2

1,2 Bauman Moscow State Technical University (Moscow, Russia)
1 bsgor@mail.ru, 2 dkkarpov@gmail.com

Today, with any experience working with large amounts of text data, a person's attention tends to dissipate, which can lead to errors. However, for example, when the teacher checks the answers to the questions of the control work, errors in evaluating the correctness of these answers are extremely undesirable. In this case, the use of automatic response verification tools will be extremely effective.

Goal – Determining the optimal set of methods and rules that allow Internet users to verify detailed answers to an abstract question, regardless of the topic and language.

The existing GPT models are considered from the point of view of checking extended answers to questions. The model was tested on a dataset with English-language questions. The principles of forming queries to large language models have been developed and substantiated.

The principles of forming queries to large language models are substantiated, which will simplify the process of checking and evaluating detailed answers to questions, as well as the formulation of principles of forming queries to large language models for the greatest optimization of work with them.

Goryachkin B.S., Karpov D.K. Development and substantiation of the principles of query formation for large language models. Science Intensive Technologies. 2024. V. 25. № 6. P. 60−68. DOI: https://doi.org/10.18127/j19998465-202406-09 (in Russian)

1. Terekhov V.I., Kanev A.I. Sistema izvlecheniya informacii iz teksta dlya metagrafovoj bazy znanij. Dinamika slozhnyh sistem – XXI vek. 2020. № 3. S. 57–66. DOI 10.18127/j19997493-202003-0 (in Russian).
2. Bol'shie yazykovye modeli (LLM). URL: https://habr.com/ru/articles/768844/ (data obrashcheniya: 30.11.2023) (in Russian).
3. 200,000+ Jeopardy! Questions. URL: https://www.kaggle.com/datasets/tunguz/200000-jeopardy-questions (data obrashcheniya: 30.11.2023).
4. Karpov D.K. Proverka tekыstovyh zadanij s pomoshch'yu bol'shih yazykovyh modelej. Tendencii razvitiya nauki i obrazovaniya. 2023.
   № 103. Ch. 8. S. 150–156. DOI 10.18411/trnio-12-2023-p8 (in Russian).
5. GPT-4 protiv GPT-3. URL: https://appmaster.io/ru/blog/gpt-4-protiv-gpt-3#:~:text=Osnovnye%20razlichiya%20mezhdu%20GPT-4,pri%20rabote%20s%20dlinnymi%20posledovatel'nostyami (data obrashcheniya: 10.12.2023) (in Russian).
6. Na chto sposobna nejroset' GPT-3. URL: https://trends.rbc.ru/trends/industry/cmrm/602c161c9a79474c3372027c (data obrashcheniya: 10.12.2023) (in Russian).
7. Istoriya nejroseti ChatGPT. URL: https://workspace.ru/blog/istoriya-chatgpt-chto-umeet-chat-gpt-4-i-chego-zhdat-ot-gpt-5/#:~:text= kolichestvo%20parametrov%20—%20175%20milliardov%3B,96%20sloev (data obrashcheniya: 7.12.2023) (in Russian).
8. GPT-4 uzhe ne za gorami. CHto my o nem znaem. URL: https://habr.com/ru/companies/cloud4y/articles/667278/ (data obrashcheniya: 30.11.2023) (in Russian).
9. Uchimsya delat' pravil'nye zaprosy dlya ispol'zovaniya v marketinge. URL: https://habr.com/ru/articles/729420/ (data obrashcheniya: 7.12.2023) (in Russian).
10. Bahman A.A., Vasyunin M.A., Galkin V.A., Gapanyuk Yu.E. Podhod k generacii tekstov programm na osnove nejrosetevyh algoritmov. Dinamika slozhnyh sistem – XXI vek. 2023. № 3. S. 58–63. DOI 10.18127/j19997493-202303-08 (in Russian).