flowstate-cua: a 155 GB dataset for fluid and heat simulation agents
flowstate-cua is a 155 GB dataset of 8,260 worked examples and 269,058 interactions for training AI agents on fluid and heat simulations in ParaView.
introducing flowstate-cua 🤗 (155 GB) a dataset for training ai agents to work with fluid and heat simulations. the 155 gb release contains 8,260 worked examples and 269,058 interactions: task instructions, screenshots, mouse and keyboard actions, simulation fields, saved sessions, and reference results. the fields are synthetic; the software interactions are recorded in paraview. use it to fine-tune models, evaluate complete workflows, or build assistants that answer questions like “what fraction of this system is overheating?” in our evaluation, a 4b model went from 0/120 to 20/120 completed tasks after fine-tuning on the scientific collection. there’s room to expand with 3d problems, more recovery scenarios, and testing across more models and simulation tools. the original data and code are mit-licensed. explore the dataset and code https://huggingface.co/datasets/harrrshall/FluidThermal-CUA