Dataset for 'A Hybrid GUI-LLM Interface Paradigm for 3D Scene Customisation'

This dataset and fine-tunning entries accompany the paper "A Hybrid GUI-LLM Interface Paradigm for3D Scene Customisation". The paper looks for empirical evidence of GUI-LLM modality trade-offs in 3D interaction. It compiles learnings from testing with 12 participants who used a GUI and LLM interface to customise a 3D scene showing a digital replica of a city. The included files are: Anonymised participant task completion data: this files includes the time it took each participants to complete 12 tasks, using different interface types. Tasks were completed in different order and with different interfaces by each participant. Time is recorded in seconds. Additionally, we tracked the number and type of "errors" done by each participant, as well as the times the LLM made a mistake and the participant spotted it. Fine-tune samples: 150 ad-hoc prompts and responses to fine tune Mistral Smnall 3.2. This uses the traditional paradigm "system, user, assistant". System instructions: This instructions are added to the LLM to improve responses. To protect the participants’ privacy, all data have been anonymised. Personally identifiable information and sensitive content have been removed or redacted prior to publication. The dataset is provided to support transparency and reproducibility of the research reported in the corresponding publication.