Abstract:
This article addresses the problem of coordinating heterogeneous software tools in heterogeneous distributed application execution environments. Here, manually configuring launch parameters for newly installed programs on a computing cluster (such as command-line switches, environment variable values, and configuration file settings) poses significant challenges for domain researchers due to the large volume of utility information and the need to store and aggregate information in a fixed format. We propose a method for the automated extraction of launch parameters based on a hybrid neural network training architecture that combines the generation of training samples using large language models with the subsequent fine-tuning of a compact transformer encoder. This approach eliminates the need for expensive graphics accelerators by applying the Low-Rank Adaptation (LoRA) technique to models with up to 1 billion parameters, enabling model execution (inference) on standard CPUs in control nodes. To formalize the quality of extraction, a two-component metric has been developed that aggregates the structural correctness of the output JSON schema (the presence of required fields and program parameter types in the obtained data) and the semantic accuracy of parameter values (correspondence with the description in the documentation). The experimental evaluation of the method focuses on a corpus of software package documentation (man pages, README files). The design results confirm the possibility of approximating the documentation analysis process with a compact model, which contributes to the automation of the software deployment lifecycle and the reduction of task flow management errors in distributed computing systems.