VERIFICATION OF THE TEXT AFTER GENERATION BY MEANS OF LARGE LANGUAGE MODELS FOR FILTRATION OF THE INCORRECT ANSWERS

Abstract

Nowadays the problem of matching large language models becomes relevant. Models are able to perform various tasks, using zero-shot approach. But as they became more intelligent, they find alternative routes for the solution of the tasks not as the researches expect. This is especially dangerous in the production environment because it is difficult to control the output of the model which was taught to be universal. In the given paper it is suggested to use one and the same model several times in different form in order to improve the quality of the generated text.
Method of accuracy increase of the models of the text content generation was further developed. In this case the user has not to provide tens of examples of the desired and non-desired behavior of the model because the model itself can do this automatically. That is, unlike the conventional methods of model accuracy enhancement, which require the training set of the models, the proposed approach includes the identification stage. As a result of identification we obtain the set of examples, on which the model learns and enhances its accuracy.
Two specific methods were proposed in the given research. The first method simply uses the model of the discriminator for the verification of the results of the generator model and repeats the request to create the text, if the results do not correspond the criteria of the user. By means of this method all the incorrect generations were removed but due to denotating the third of the correct generations as incorrect. The second approach is more complicated, besides the discriminator it uses the model of the simulator. The process required that the simulator model generated the samples of the data, entered by the user, after that the generator would generate the text of the answer for each sample and the discriminator would verify the generated results and add them to training data. This will increase the accuracy from 56 % to 66 % in the problem of logic conclusion.