README.md

metadata

dataset_info:
  features:
    - name: prompt
      dtype: string
    - name: tool
      dtype: string
    - name: args
      struct:
        - name: a
          dtype: float64
        - name: b
          dtype: float64
  splits:
    - name: train
      num_bytes: 2617582
      num_examples: 20000
  download_size: 999433
  dataset_size: 2617582
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
tags:
  - function-calling
  - agent
  - synthetic
  - mathematics
  - nlp

Simple Calculator: Agentic Function-Calling Dataset

Overview

This dataset contains 20,000 highly structured, synthetically generated examples designed to train Large Language Models (LLMs) in agentic tool-calling and parameter extraction. It focuses on mapping diverse natural language mathematical queries—ranging from direct commands to conversational word problems—into deterministic, executable JSON structures for a calculator tool.

Dataset Structure

The dataset trains models to isolate numerical entities and identify the correct mathematical operation required to solve the user's prompt.

Schema

• prompt (string): The natural language input (e.g., "If I have a 200 bill and I want to split it with 3 friends").
• tool (string): The target target routing function (add, subtract, multiply, or divide).
• args (struct): The extracted numerical parameters.
  • a (float64): The first numerical operand.
  • b (float64): The second numerical operand.

Data Example

{
  "prompt": "A factory produces 45.5 items per hour. How many items does it make in 8 hours?",
  "tool": "multiply",
  "args": {
    "a": 45.5,
    "b": 8.0
  }
}