erfanSO/financial-fundamentals-tool-calling

Financial Fundamentals Tool-Calling Dataset

This dataset contains synthetic examples for training and evaluating financial tool-calling models. Each example pairs a natural-language user request about company fundamentals with a structured JSON function call.

The dataset is designed for models that should translate financial requests into tool calls instead of answering directly.

Task

Given a user query such as:

Retrieve net income and diluted shares for Broadcom from 2006 to 2018.

The model should produce a JSON call like:

{
  "action": "call",
  "function": "get_fundamentals",
  "arguments": {
    "queries": [
      {
        "symbols": ["Broadcom"],
        "metrics": ["Net Income", "Diluted Shares"],
        "start_year": 2006,
        "end_year": 2018
      }
    ]
  }
}

Dataset Fields

The dataset includes fields such as:

• —query: The natural-language financial request.
• —completion: The target JSON tool call.
• —text: Chat-formatted training text.
• —messages: Chat-style system, user, and assistant messages.
• —metadata: Generation metadata, including companies, metrics, years, and grouping information.
• —source_index: Index of the original source example.

Use Cases

This dataset can be used to train or evaluate models for:

• —Financial function calling
• —Structured JSON generation
• —Tool routing for agentic finance systems
• —Fundamentals retrieval request parsing
• —Company-metric-year extraction

Limitations

The dataset is synthetic and focused on financial fundamentals requests. It does not contain live financial values, investment advice, or real-time market data. Models trained on this dataset should have their JSON outputs validated before any downstream tool execution.