iselabvn/Linux-terminal-tool-calling

Linux Terminal Tool Calling Dataset (Linux-terminal-tool-calling)

This dataset is designed for training and fine-tuning AI agents on tool calling, reasoning, and command execution specifically for standard Linux terminal utilities and system administration tasks. It transforms raw Linux terminal command records into a structured multi-turn conversation format featuring detailed chain-of-thought/reasoning content and OpenAI/OpenClaw-style function calling.

Dataset Details

• —Total Records: 600
• —Language: English
• —Format: JSONL (JSON Lines)
• —License: Apache 2.0
• —Repository: iselabvn/Linux-terminal-tool-calling

Dataset Structure

Each record is formatted as a single-turn conversation with user and assistant roles, complemented by rich metadata for downstream filtering and analysis.

Field Descriptions

• —`messages` (list): Conversation history.
• —`role: "user"` (dict): The user request describing a Linux terminal or system administration task.
• —`role: "assistant"` (dict): The assistant response containing:
• —`reasoning_content` (str): A 2-3 sentence chain-of-thought explanation explaining the choice of command, flags, and parameter configurations.
• —`tool_calls` (list): An array containing the function call. The tool utilizes the exec function to run the command on the target environment.
• —`id` (str): A unique call identifier (e.g., call_exec_0).
• —`type: "function"` (str): The type of tool call.
• —`function` (dict): Details of the target function call.
• —`name`: "exec"
• —`arguments` (JSON string): Serialized dictionary containing the exact executable command ({"command": "..."}).
• —`content` (str): Empty string (standard format for tool calling responses).
• —`metadata` (dict): Metadata associated with the command execution.
• —`id` (str): Unique command ID (e.g., cmd-001).
• —`category` (str): Functional category of the command (e.g., File Management, Viewing, System Info).
• —`example_output` (str): Sample execution output or return status.
• —`man_reference` (str): Link to the official man page documentation.
• —`original_description` (str): The original, generic user description before prompt synchronization.

Category Distribution

Sample Record

{
  "messages": [
    {
      "role": "user",
      "content": "Change the current working directory to /var/log."
    },
    {
      "role": "assistant",
      "reasoning_content": "The user wants to navigate to the /var/log directory. The 'cd' command is the standard way to change directories in Unix-like systems. The path /var/log is specified directly. Executing the tool.",
      "tool_calls": [
        {
          "id": "call_exec_0",
          "type": "function",
          "function": {
            "name": "exec",
            "arguments": "{\"command\": \"cd /var/log\"}"
          }
        }
      ],
      "content": ""
    }
  ],
  "metadata": {
    "id": "cmd-001",
    "category": "Navigation",
    "example_output": "[No output, changes directory]",
    "man_reference": "https://man7.org/linux/man-pages/man1/cd.1.html",
    "original_description": "Changes the current directory to /var/log."
  }
}

Intended Use

This dataset is ideal for:

• —Fine-tuning language models to act as autonomous agent loops in Linux terminal environments.
• —Training models on standard OpenAI tool calling conventions for shell commands.
• —Supervised Fine Tuning (SFT) for system administration assistants, incorporating chain-of-thought (reasoning) before issuing commands.

Construction Method

The dataset was constructed by converting raw Linux terminal command records into a structured tool-use conversation trace.

To ensure consistency between user requests and executable commands, we utilized internal LLMs to perform Prompt Synchronization:

1. 1.Target Injection: Generic references (e.g., "a directory", "a file") in user requests were automatically replaced or synchronized with specific target parameters found in the command (e.g., /var/log, logfile.txt).
2. 2.Chain-of-Thought Synthesis: The LLM generated a 2-3 sentence reasoning_content to justify the selection of the command, flags, and arguments.
3. 3.Metadata Preservation: Original command IDs, categories, man references, sample outputs, and descriptions are preserved in metadata for alignment and verification.