Skip to main content

Explanation

Data Preparation​

You can run ScienceAgent on your own data by preparing a CSV file. Each entry should have at least the following columns:

  • task_inst: task instruction
  • domain_knowledge: optional domain knowledge, usually annotated by subject matter experts.
  • dataset_folder_tree: folder tree structure of the dataset(s)
  • dataset_preview: preview of a few entries in the dataset(s)

We provide an example task from ScienceAgentBench in benchmark/ScienceAgentBench.csv.The additional fields are for evaluation of agent performance, which is not covered by this repository. For a full evaluation on ScienceAgentBench, please refer to the ScienceAgentBench repo!