Vladyslav Kotsovsky

This study examines the quality of SQL queries in Text-to-SQL datasets, focusing on antipatterns in gold-standard queries used to train and evaluate large language models (LLMs). While such queries may be syntactically valid and executable, they can still contain structural flaws that affect correctness, portability, and benchmark reliability. Since these flaws do not prevent successful execution, they often go undetected by conventional validation methods — creating what we call a correctness paradox. A rule-based static analysis framework was designed and implemented for detecting 14 SQL antipatterns in Text-to-SQL datasets. The rules operate on abstract syntax trees produced by a dialect-aware SQL parser. Each detected antipattern is classified into one of four severity levels (Critical, High, Medium, Low) using a configurable profile for SQLite, PostgreSQL, and other RDBMSs. The framework was applied to six partitions from three major Text-to-SQL benchmarks: Spider (11,840 queries), BIRD (10,962 queries), and the Gretel synthetic dataset (100,000 queries). Overall, 82.5–89.7% of queries were antipattern-free, and 16,554 antipattern occurrences across 14 types were detected. Spider was dominated by Missing GROUP BY, BIRD by Function in WHERE, whereas Gretel exhibited the broadest defect spectrum and was the only dataset containing 16 syntactically invalid SQL queries. Manual verification of all Critical-severity cases confirmed that the vast majority represented genuine defects; a small number of borderline cases were retained due to the framework's deliberately conservative detection policy. A comparative evaluation on a representative sample of 1,237 queries showed that, while our framework detected antipatterns across all 14 classes, the competing tool identified the expected antipattern in only 58.6% of cases, covering just 4 classes. For Cartesian products specifically, the competing tool reached only 29.5% recall, missing eight distinct syntactic patterns. Based on these findings, we recommend integrating automated antipattern detection into benchmark quality assurance workflows, prioritizing and fixing Critical-severity cases across the analysed datasets, correcting the 16 syntactically invalid SQL queries identified in the Gretel dataset, and resolving Missing GROUP BY patterns that undermine cross-engine portability. To support adoption of these recommendations, the framework is released as an open-source tool.

The paper treats the application of the bithreshold approach in the design of neural network classifiers. A novel hybrid 3-layer neural network model is proposed whose first hidden layer consists of bithreshold neurons, and the second hidden layer employs the softmax activation function. This model is intended to solve multiclass classification tasks. A supervised synthesis algorithm is designed for this neural network architecture. It consists of two stages. During the first stage a given training pattern is separated from representatives of other classes using single-threshold neurons, which are gradually converted into bithreshold neural units. In the second stage, the network design procedure reduces the size of network hidden layers in order to simplify the network and enhance the recognition ability. The performance of the proposed model is compared with that of several popular machine learning classifiers on a real-world dataset. Simulation results on optical recognition of handwritten digits benchmark demonstrate that the developed neural network model is suitable for multiclass classification tasks.

The modification of the online learning algorithm for multi-valued multithreshold neurons is proposed in the paper. Conditions are stated and proved that ensure the finite successful learning. The influence of the algorithm hyperparameters on the learning process is analyzed on the base of simulation results. The recommendations are formulated concerning the choice of values of these hyperparameters, which may significantly reduce the learning time. The experiment results prove that the proposed algorithm and Parberry. Obtained results can be useful in the design of artificial neural network classifiers employing multithreshold activation functions in network nodes.

The ways to improve the performance of classifiers based on the bithreshold architecture are considered in the paper. The model of 3-layer neural network binary classifier is proposed whose first hidden layer consists of neural units of 3 kinds: bithreshold neurons, linear threshold units and winner-take-all neurons, and every neuron in the second layer has only 3 inputs with predefined weights. The synthesis algorithm for such networks is designed and estimations of its time complexity and the size of resulting network are presented. The simulation results demonstrate that the application of the new architecture in the classifier design significantly improves its generalization ability.

The fully connected 2-layer feedforward network architecture with the hidden layer consisting of bithreshold neurons is considered in the paper. We design the neural network multicategory classifier based on such an architecture. We also give the upper bound on the size of the bithreshold network, which is capable to recognize an arbitrary partition of the given finite set of patterns into a predefined number of classes. The learning algorithm is proposed for such neural networks. We give simulation results of the classifier performance on the real data set and discuss the influence of algorithm parameters on the quality of learning and the representational capacity of classifier.