Abstract
Novel large pre-trained language models, such as GPT-3, can be considered and adopted as artificial agents since they are now able to solve general problems and mimic human experts. The introduction of in-context learning technique opens up the possibility of interacting with the model directly by instructing it to solve a task. Task instructions, the actual input data, and optionally some examples of solutions are packed together in a single prompt. The model interprets the prompt and generates a solution for the given problem without any need of fine-tuning the model. However, designing efficient prompts is more an art than a science, nowadays. When starting from scratch, to achieve good performance different prompt contents and model engine’s configurations (for the same prompt) must be tested. These can be considered time-intensive operations. In this paper we present Experiment Maker a software developed to save time and minimize the effort in designing and testing different prompts and different configurations, In addition, the tool supports users in combining multiple prompts into an experimental pipeline. Experiment Maker can be downloaded from the project page at github.com/patriziobellan86/ExperimentMaker.