Parameterizing LLMs in Practice: An Empirical Study of LLMs Integrated into Software Systems

This dataset contains information on 565 open-source software projects from GitHub that integrate Large Language Models. The projects were selected based on: using the Python programming language having at least three defined topics from the following: 'ai', 'chatgpt', 'dall-e', 'generative-ai', 'generativeai', 'gpt', 'llm', 'chatbot', 'transformers', 'language-model', 'chatbots', 'llms', 'agent', 'gemini', 'llama', 'app', 'ui', 'claude', 'artificial-intelligence', 'large-language-models', 'lora', 'mistral', 'langchain', 'assistant', 'ollama', 'agents', 'bert'. having at least 100 stars For each of these projects, the definitions of the LLM parameters used in the code were extracted, namely 'temperature', 'top-k', 'top-p', 'min-p', 'repetition-penalty', 'frequency-penalty', 'presence-penalty', and 'max-tokens'. The final dataset contains 7,892 LLM parameter definitions extracted from 363 projects. It should be noted that when extracting the parameters from the 565 selected projects, solely 363 projects remained because there are projects where we have not found defined parameters that are used in the production code (i.e., not taking into account test files, documentation, examples, etc.).