Russian parametric corpus RuParam

The main function of large language models is to simulate the behavior of native speakers in the most correct way. Hence, it is necessary to have assessment datasets to track progress in solving this problem as well as regularly compare competing models with each other. There are some datasets of this type, the so-called linguistic acceptability corpora. The hypothesis that underlies the creation of these corpora assumes that large language models, like native speakers, should be able to distinguish correct, grammatical sentences from the ungrammatical ones that violate the grammar of the target language. The paper presents the parametric corpus for Russian, RuParam. Our corpus contains 9.5 thousand minimal pairs of sentences that differ in grammaticality — each correct sentence corresponds to a minimally different erroneous one. The source of ungrammaticality in each pair is supplied with the linguistic markup provided by experts. RuParam consists of two parts: the first part uses a totally new data source for the task of testing large language models — lexical and grammatical tests on Russian as a foreign language. The second part consists of (modified and tagged) examples from real texts that represent grammatical phenomena, not included in the RFL teaching program due to their complexity. As have shown our experiments with different Large Language Models, the highest results are achieved by those models that have been trained on Russian most carefully at all stages, from data preparation and tokenization to writing instructions and reinforcement learning (these are first of all YandexGPT and GigaChat). Multilingual models, which usually receive little or no emphasis on Russian, showed significantly lower results. Still, even the best models results are far from the assessors who completed the task with almost 100 % accuracy. The models ranking obtained during the experiment shows that our corpus reflects actual degree of proficiency in Russian. The resulting rating can be helpful when choosing a model for natural language processing task requiring grammar knowledge: for example, building morphological and syntactic parsers. Thereafter, the proposed corpus can be used to test your own models.