Domain_Specific_Financial_LLM

Domain-specific LLM system for financial risk analysis and document-based question answering, combining fine-tuned finance models deployed via Ollama with a lightweight Streamlit application and a supporting RAG + finance data layer.

• Main application: stream_app.py
• Fine-tuning dataset: https://huggingface.co/datasets/adityamavle/FinRiskAnalysis

---

Overview

This project implements a unified LLM application for multiple finance-specific NLP tasks, including risk analysis, sentiment analysis, named entity recognition, and document Q&A.
Models are fine-tuned on a curated financial dataset, registered with Ollama, and invoked through task-specific system prompts from a Streamlit UI.

Supported tasks:

• Risk Analysis
• Financial Sentiment Analysis
• Financial NER
• Financial Visual Data Analysis
• DocQA (document upload + retrieval-augmented QA)

---

Architecture

1. Model Fine-Tuning

• Curated and fine-tuned domain-specific LLMs on the FinRiskAnalysis dataset.
• Task-specialized models (e.g., risk, sentiment, NER) uploaded and served via Ollama.

2. Application Layer

• stream_app.py provides a chat-style Streamlit interface.
• Requests are routed to the appropriate Ollama model based on the selected task.
• Responses are streamed token-by-token for interactive usage.

3. RAG and Finance Data Layer

• DocQA uses a supporting RAG pipeline for grounded answers over uploaded documents.
• A finance data layer handles document ingestion, preprocessing, and dataset access.
• RAG logic and data handling are kept modular and separate from UI code.

---

Repository Structure

.
├── README.md
├── stream_app.py            # Main Streamlit application
├── doc.py                   # Document upload + DocQA entry point
├── rag_*                    # Retrieval-Augmented Generation utilities
├── finance_data_*           # Finance dataset loading / preprocessing

---

How to Run

Install dependencies, then run the Streamlit application:

streamlit run stream_app.py

---

Dataset

Fine-tuning dataset used in this project:

FinRiskAnalysis (Hugging Face)
https://huggingface.co/datasets/adityamavle/FinRiskAnalysis

---

Notes

• Task-specific model names (e.g., mistral-risk, fin-sentiment, mistral-NER) refer to fine-tuned Ollama models used by the app.
• Commit messages and timestamps in the repository are not part of the project description.
• The Streamlit app maintains conversational state across tasks using st.session_state.

---

Links

• Dataset: https://huggingface.co/datasets/adityamavle/FinRiskAnalysis