Artificial-Production-Units/agent-data-collection

Dataset Card for OpenHands Agent Logs

This dataset consists of multi-turn dialogues between a simulated human and an LLM-based agent interacting in a virtual operating system environment. Each conversation involves the agent reasoning about and solving command-line tasks through execute_bash and related actions.

Dataset Format

Each file in the dataset is a .json file, structured as a list of instances. Each instance contains:

• id: A unique identifier for the interaction session.
• system: System prompt defining rules, roles, and constraints for the agent.
• conversations: A list of alternating human and gpt turns, with structured interactions including thought steps, tool use, and solutions.

Example

{
  "id": "os_0",
  "system": "You are OpenHands agent...",
  "conversations": [
    {
      "from": "human",
      "value": "tell me how many files are in the directory \"/etc\"?"
    },
    {
      "from": "gpt",
      "value": "<function=execute_bash>..."
    },
    ...
  ]
}