Finance Chatbot: Sentiment Analysis & Investment Advisor

Author: Atharva Kulkarni

Key Features

• Sentiment Analysis of financial news/text using fine-tuned LLaMA-3
• Investment Recommendations with real-time stock metrics integration
• Financial Q&A system with contextual understanding
• LoRA-optimized fine-tuning for efficient adaptation
• Real-time data integration with Yahoo Finance & NewsAPI

Technical Stack

| Component | Technology | |————————-|————————————-| | Base Model | Meta-Llama-3-8B-Instruct | | Fine-tuning Framework | Unsloth | | Quantization | 4-bit QLoRA | | Data Processing | Pandas, Hugging Face Datasets | | Real-time Data | yfinance, NewsAPI | | Deployment | Transformers, PEFT |

Installation

# Clone repository
git clone https://github.com/Atharvack/LLM-Finetuning-Yfinance.git
cd finance-chatbot

# Additional setup
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install --no-deps xformers trl peft accelerate bitsandbytes

Usage

Fine-tuning Process

from unsloth import FastLanguageModel

# Initialize model with LoRA
model, tokenizer = FastLanguageModel.from_pretrained(
  model_name="meta-llama/Meta-Llama-3-8B-Instruct",
  max_seq_length=1024,
  load_in_4bit=True,
  token="your_hf_token"
)

# Start training
trainer = SFTTrainer(
  model=model,
  train_dataset=dataset,
  dataset_text_field="text",
  max_seq_length=1024,
  # ... additional config
)
trainer.train()

Real-time Inference

# Initialize chatbot interface
def main():
  context = build_context()  # Collects real-time data
  while True:
    question = input("Ask financial question: ")
    response = inference(question, context)
    print(f"Assistant: {extract_response(response)}")

📊 System Architecture

graph TD A[User Input] --> B[Data Collection Module] B --> C[yFinance API] B --> D[NewsAPI] C --> E[Real-time Metrics] D --> F[News Context] E --> G[LLaMA-3 Processor] F --> G G --> H[Sentiment Analysis] G --> I[Investment Advice] H --> J[Response Generation] I --> J J --> K[Output]

Dataset

• QA_plus_sentiment_12k.csv containing:
  • Financial questions
  • Contextual information
  • Annotated answers
  • Sentiment labels

Configuration

1. Obtain Hugging Face Token
2. Get NewsAPI Key
3. Update hf_token and api_key in code

License

MIT License - See LICENSE for details

---

Disclaimer: This project is for educational purposes only. Never make actual financial decisions based on AI outputs without consulting qualified professionals.

Jupyter Notebooks

Check repo here

Acknowledgements

The foundational structure for the LLaMA 3 fine-tuning process was adapted from popular open-source tutorials and best practices demonstrated by the AI/ML community. Special thanks to the creators of the Unsloth library for their high-performance tools and to educators like Ashish Patel for their clear and effective guides. The core real-time data pipeline and financial application logic are original work for this project.

Try out the LLM on Huggingface [here][https://huggingface.co/spaces/Rashtardeep/Finance-llm]