pranavvmurthy26/synthetic-financial-tool-calling-grpo-rlvr-1k

🤖 Synthetic Financial Tool Calling Dataset for GRPO and RLVR

This is a synthetic dataset designed for training language models on financial tool calling using GRPO (Group Relative Policy Optimization) with verifiable rewards (RLVR). The dataset contains ~1.1K examples of financial planning queries paired with expected tool calls and answers.

Dataset sample schema,

{
  "prompt": [
    {
      "role": "system",
      "content": "You are a financial planning assistant with tools for portfolio allocation, mortgage affordability, tax optimization, retirement readiness, debt payoff strategies, insurance needs, education funding, and currency exchange.\nAnalyze user requests and call the appropriate tool with all required parameters extracted from their query.\nReturn concise answers with key metrics. Do not ask for clarification - use reasonable defaults if needed."
    },
    {
      "role": "user",
      "content": "Age 30, retiring at 58, $95k saved, $15k annual contributions, 75% employer match. Aggressive 8.5% returns, 2.5% inflation. Want $60k retirement income, Social Security $20k at 62, no pension, healthcare $9k/year. Living to 90."
    }
  ],
  "answer": "Projected Savings: $3,656,004, Annual Gap: $117,758, Readiness: 97%, Status: Nearly There",
  "ground_truth": {
    "name": "calculate_retirement_readiness",
    "arguments": {
      "current_age": 30,
      "retirement_age": 58,
      "current_savings": 95000,
      "annual_contribution": 15000,
      "employer_match_percent": 0.75,
      "expected_return": 0.085,
      "inflation_rate": 0.025,
      "desired_retirement_income": 60000,
      "social_security_estimate": 20000,
      "pension_income": 0,
      "healthcare_cost_annual": 9000,
      "life_expectancy": 90
    }
  }
}

Fine-tuning an LLM using Reinforcement learning leverages prompt and uses answer optionally to verify reward.

📄 Schema

Column	Type	Description
prompt	list[dict]	A conversation-style prompt with role (system/user) and content fields. The system message defines the assistant's capabilities, and the user message contains a natural language financial query.
answer	string	The expected human-readable output from executing the correct tool call (e.g., insurance coverage recommendations with premiums).
ground_truth	string	A JSON string containing the exact tool call specification with name (function name) and arguments (parameter dictionary) that should be invoked to answer the query.

📊 How the Data is Used

The dataset is used with a GRPO trainer for tool-calling optimization:

1. Dataset Loading: Loads prompts, answers, and ground truth tool calls
2. Model Generation: The model generates completions (tool calls) given the prompts
3. Tool Execution: Generated tool calls are executed against actual tool functions
4. Reward Computation: The reward function compares tool execution results against expected answers
5. Policy Optimization: GRPO uses rewards to optimize tool-calling behavior through relative comparisons across generations

🛠️ Tool Description and Rewards

Tools

Reference: financial_tools_complex.py

The dataset targets 8 financial planning functions with complex argument structures (8-12 parameters each):

Tool	Purpose
calculate_portfolio_allocation	Optimal asset allocation based on risk tolerance, time horizon, age, and preferences
calculate_mortgage_affordability	Maximum home price and monthly payment breakdown
optimize_tax_strategy	Tax liability calculation with optimization recommendations
calculate_retirement_readiness	Retirement savings projection and readiness assessment
analyze_debt_payoff_strategy	Optimal debt payoff approach (avalanche/snowball/hybrid)
calculate_insurance_needs	Life, disability, and LTC insurance coverage recommendations
calculate_education_funding	College savings projections and funding gap analysis
calculate_currency_exchange_arbitrage	Currency exchange optimization with fee analysis

Each function returns a deterministic string result, enabling exact-match reward computation during training.

📈 Reward Function

Reference: financial_tools_reward.py

The reward function implements a 3-tier scheme:

Reward	Condition
1.0	Tool response exactly matches the expected answer
0.1	A tool was called but the response doesn't match
0.0	No tool call was made

This structure encourages the model to: (1) learn to make tool calls, and (2) learn to make the correct tool calls with proper arguments.

🦾 Dataset generation

This dataset was generated with the help of kiro.dev.

Citation

If you use this dataset, please cite:

@dataset{pranavvmurthy26-synthetic-financial-tool-calling-grpo-rlvr-1k,
  title={Synthetic Financial Tool Calling GRPO RLVR Dataset},
  author={Murthy, Pranav},
  journel={},
  year={2026}
}