Gandalf1/indian-finance-synthetic-phase2

Indian Finance Synthetic Dataset - Phase 2

A high-quality synthetic dataset of 14,835 Indian personal finance conversations for fine-tuning language models.

Dataset Description

This dataset contains synthetic conversations between users seeking personal finance advice and a financial assistant (FinEdge). All conversations are tailored to the Indian context, covering tax planning, investments, insurance, goal planning, and more, based on FY 2024-25 regulations.

Key Features

• —Size: 14,835 samples
• —Format: OpenAI messages format (system, user, assistant)
• —Quality: Heavily filtered (26.2% rejection rate) to remove truncated, incomplete, or low-quality samples
• —Generation: Created using DeepSeek v4 (generation) + GPT-4o-mini (8-dimension evaluation)
• —Domain: Indian personal finance with tools integration
• —Language: Primarily English with Hindi terms and Indian financial terminology

Dataset Statistics

Length Distribution:

• —Average user query: 373 characters
• —Average assistant response: 1,696 characters
• —Average total conversation: 4,097 characters
• —Query range: 27 - 1,234 characters
• —Response range: 364 - 10,674 characters

Conversation Types:

• —Tool required: 35.2% (5,216 samples)
• —Conceptual (no tool): 30.5% (4,525 samples)
• —Clarification needed: 21.7% (3,221 samples)
• —Multi-tool reasoning: 7.6% (1,134 samples)
• —Edge cases: 5.0% (739 samples)

Topics:

• —Goal planning: 37.7%
• —Tax planning: 31.8%
• —Conceptual: 30.5%

Difficulty Levels:

• —Intermediate: 68.2%
• —Advanced: 29.0%
• —Beginner: 2.8%

Data Structure

Each sample contains:

{
  "messages": [
    {
      "role": "system",
      "content": "System prompt with FinEdge persona and instructions"
    },
    {
      "role": "user",
      "content": "User's financial question or scenario"
    },
    {
      "role": "assistant",
      "content": "Assistant's response with <think> tags and advice"
    }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "calculate_goal_sip",
        "description": "...",
        "parameters": {...}
      }
    }
  ],
  "metadata": {
    "conversation_type": "tool_required",
    "topic": "goal_planning",
    "difficulty": "intermediate",
    "reasoning_style": "step_by_step",
    "reasoning_depth": "detailed"
  }
}

Quality Assurance

Filtering Applied:

• —✅ Removed empty/missing user messages (3,341)
• —✅ Removed queries < 20 characters (89)
• —✅ Removed duplicates
• —✅ Removed queries ending with commas (66)
• —✅ Removed responses containing "incomplete" (72)
• —✅ Removed dash endings (22)
• —✅ Removed other special character endings (47)
• —✅ Removed truncated queries ending with letters/numbers (1,619)

Total samples removed: 5,256 (26.2%) Final clean dataset: 14,835 samples

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Gandalf1/indian-finance-synthetic-phase2")

Fine-tuning Example

This dataset is formatted for OpenAI-compatible fine-tuning:

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load dataset
dataset = load_dataset("Gandalf1/indian-finance-synthetic-phase2")

# Use for fine-tuning
# Your fine-tuning code here

Model Recommendations

This dataset is suitable for:

• —8B parameter models (optimal size)
• —LLaMA 3, Mistral, Qwen, Gemma families
• —Function calling / tool use training
• —Indian finance domain adaptation

Limitations

• —Synthetic data generated by LLMs (DeepSeek v4)
• —May contain biases from generation model
• —Focused on Indian context (FY 2024-25)
• —Not a replacement for professional financial advice
• —Should be validated before production use

Citation

If you use this dataset, please cite:

@dataset{indian_finance_synthetic_phase2,
  author = {Yanamandra, SA},
  title = {Indian Finance Synthetic Dataset - Phase 2},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Gandalf1/indian-finance-synthetic-phase2}
}

License

Apache 2.0

Contact

For questions or issues, please open an issue on the dataset repository.