finance-chat-assistant

Finance Chat Assistant

Model Description

Finance Chat Assistant is a finance-focused conversational AI model created by fine-tuning TinyLlama-1.1B-Chat-v1.0 using QLoRA (Quantized Low-Rank Adaptation) on a finance instruction dataset.

The objective of this project is to improve the model's ability to answer finance-related questions including:

• Diversification
• Mutual Funds
• ETFs
• Asset Allocation
• Inflation
• Risk Management
• Retirement Planning
• Investment Strategies
• SIP Investing

This repository contains LoRA adapter weights only and requires the TinyLlama-1.1B-Chat-v1.0 base model for inference.

---

Model Details

Base Model

TinyLlama/TinyLlama-1.1B-Chat-v1.0

Fine-Tuning Method

• QLoRA
• PEFT
• LoRA
• TRL SFTTrainer
• 4-bit NF4 Quantization

Author

Kuran (GitHub: kurannnnn)

---

Demo

A Streamlit-based Finance Chat Assistant application was developed using this model.

Features:

• Finance question answering
• Investment concept explanations
• Educational finance assistant
• Interactive chat interface

---

Model Sources

Hugging Face Repository

https://huggingface.co/Konnnnnnnn/finance-chat-assistant

Base Model

https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0

Project Repository

https://github.com/kurannnnn/Finance-Chat-Assistant

---

Uses

Direct Use

This model can be used for:

• Finance education
• Investment concept explanations
• Personal finance learning
• Finance chatbots
• Academic projects involving LLM fine-tuning

Downstream Use

The model can be integrated into:

• Streamlit applications
• Educational finance assistants
• Learning platforms
• RAG systems
• Research projects

Out-of-Scope Use

This model should not be used for:

• Professional financial advice
• Tax advice
• Legal advice
• Real-money investment decisions
• Financial planning without human verification

---

Bias, Risks and Limitations

This model was trained on a limited finance instruction dataset and may:

• Produce inaccurate financial information
• Generate outdated recommendations
• Hallucinate facts
• Provide incomplete investment guidance

Users should verify financial information from trusted sources before making decisions.

Recommendation

This model is intended for educational and research purposes only.

---

Example

Input

What is asset allocation in a portfolio?

Output

Asset allocation refers to the distribution of investments across different asset classes such as stocks, bonds, and cash. The goal is to balance risk and return according to an investor's objectives, financial goals, and risk tolerance.

---

How to Get Started

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
)

tokenizer = AutoTokenizer.from_pretrained(
    "Konnnnnnnn/finance-chat-assistant"
)

model = PeftModel.from_pretrained(
    base_model,
    "Konnnnnnnn/finance-chat-assistant"
)

---

Training Details

Dataset

Finance instruction dataset containing question-answer pairs covering:

• Personal Finance
• Mutual Funds
• SIP Investing
• ETFs
• Asset Allocation
• Diversification
• Inflation
• Risk Management
• Retirement Planning

Training Summary

Trainable Parameters

Only 0.1023% of the model parameters were trained using LoRA adapters, significantly reducing memory requirements and training cost compared to full fine-tuning.

---

Evaluation

Validation Loss

Best Validation Loss:

2.409989

Results

The model demonstrates domain adaptation toward finance-related conversations.

While the model produces more finance-focused responses in some cases, performance gains over the base TinyLlama model vary depending on question complexity and prompt format.

Further improvements may require:

• Larger finance datasets
• Additional training epochs
• More comprehensive evaluation benchmarks
• Human evaluation studies

---

Technical Specifications

Architecture

TinyLlama-1.1B-Chat-v1.0

Objective

Causal Language Modeling

Fine-Tuning Method

QLoRA

Adapter Format

PEFT LoRA Adapter

Model Size

• Base Model: 1.1 Billion Parameters
• Adapter Size: ~4.5 MB
• Quantization: 4-bit NF4

---

Future Improvements

• Larger finance datasets
• Better evaluation benchmarks
• RAG integration
• Human preference evaluation
• Hugging Face Spaces deployment
• Improved response quality and factual accuracy

---