FinAI-Llama3-Ko
🚀 Korean Financial Recommendation LLM fine-tuned with LoRA and enhanced with customer/product retrieval-based RAG.
Model Overview
FinAI-Llama3-Ko is a Korean financial recommendation model fine-tuned from Llama-3-Open-Ko-8B using LoRA (Low-Rank Adaptation).
The model is designed to provide personalized financial and insurance recommendations by analyzing customer profiles, investment tendencies, and financial goals.
This repository contains LoRA adapter weights only.
To use this model, load the adapter together with the base model:
Base Model: beomi/Llama-3-Open-Ko-8B
Key Features
Personalized Financial Consultation
The model generates recommendations based on:
- Age
- Gender
- Income level
- Occupation
- Credit rating
- Risk preference
- Existing financial products
- Financial goals
Explainable Recommendation Generation
Instead of recommending products directly, the model is trained to generate:
- Customer Analysis
- Investment Profile Analysis
- Recommendation Rationale
- Final Recommendation
This enables more transparent and explainable financial advice.
Insurance & Securities Support
Supported domains:
- Insurance products
- Securities products
- Retirement planning
- Savings planning
- Investment recommendations
- Lifecycle-based financial consultation
Dataset
This model was fine-tuned using the labeled dataset from the AI Hub project:
Financial Products, Services and Consumer Characteristics Dataset
- Provider: AI Hub (Korea)
- Language: Korean
- Domain: Finance / Insurance / Securities
Training Samples
| Category | Samples |
|---|
| Securities | 6,400 |
| Insurance | 1,600 |
| Total | 8,000 |
The dataset contains:
- Customer profiles
- Financial product information
- Investment preferences
- Recommendation rationales
- Consultation scenarios
Data Processing
The original structured consultation data was transformed into an instruction-following format suitable for Supervised Fine-Tuning (SFT).
Training examples follow the format:
1<s>[INST]
2당신은 생애주기 기반 금융/보험 전문 AI 어드바이저입니다.
3
4[분야]
5보험 또는 증권
6
7[고객질문]
8...
9[/INST]
10
11[고객분석]
12...
13
14[투자성향]
15...
16
17[추천근거]
18...
19
20[최종답변]
21...
22</s>
Reasoning structure:
1고객질문
2 ↓
3고객분석
4 ↓
5투자성향
6 ↓
7추천근거
8 ↓
9최종답변
Training Configuration
| Parameter | Value |
|---|
| Base Model | beomi/Llama-3-Open-Ko-8B |
| Fine-tuning Method | LoRA |
| Epochs | 2 |
| Learning Rate | 2e-4 |
| Batch Size | 1 |
| Max Sequence Length | 2048 |
| LoRA Rank (r) | 16 |
| LoRA Alpha | 32 |
| GPU | NVIDIA A100 |
| Framework | Transformers + PEFT |
Example
Input
1[분야]
2증권
3
4[고객질문]
5현재 자산은 약 4천만원이며, 결혼자금 마련이 목표입니다. 투자기간은 5년 이상이고, 원금 손실은 크지 않았으면 좋겠습니다.
Output
1고객님께서는 결혼자금 마련이라는 명확한 목적과 함께, 비교적 짧은 투자 기간(5년 이상)을 갖고 계시며, 큰 원금손실에는 부담을 느끼시는 것으로 파악됩니다.
2이에 따라 단기적인 변동성이 낮고 중도 환매가 가능한 유동성을 중요하게 고려하셨으며, 세제혜택이나 현금흐름에도 관심을 두고 있습니다.
3'OOOOOO OO'는 국공채 등 우량 채권에 투자해 위험이 매우 낮고, 평균 1.62%의 과거 성과를 보여왔습니다.
4'OOO OOOOOOO' 역시 만기가 없고 언제 든 해지가 가능하며, 신용도가 높은 기업어음 중심으로 운용되어 낮은 위험을 추구합니다.
5두 상품 모두 위험등급이 6등급(매우낮은위험)으로 분류되어 있어, 고객님의 요구 조건인 원금 손실 최소화와 자금 활용 유연성을 동시에 만족할 수 있는 선택지라고 할 수 있겠습니다.
6다만 각 상품마다 세제혜택 제공 여부 또는 추가 납입 방식, 그리고 실제 투자 대상 차이가 있을 수 있으므로, 구체적인 장단점까지 비교하신 후 결정하시면 더욱 좋겠습니다.
RAG Pipeline (Implemented)
FinAI-Llama3-Ko incorporates a RAG-based recommendation system.
- Customer Vector Database (335,506 customer records)
- Product Vector Database
- FAISS-based Semantic Search
- Similar Customer Retrieval
- Personalized Recommendation Generation
Future Work
Intended Use
Suitable for:
- Financial recommendation research
- Korean financial NLP research
- Financial chatbot development
- Insurance recommendation systems
- Retrieval-Augmented Generation experiments
Not intended for:
- Real financial advice
- Investment guarantees
- Production financial services without human review
Limitations
- Recommendations are generated from historical training data.
- Financial products and regulations may change over time.
- Outputs should be reviewed by qualified financial professionals before use in real-world financial decision-making.
Disclaimer
This model is intended for research and educational purposes only.
The generated outputs should not be interpreted as professional financial, investment, insurance, or legal advice.
Users are responsible for verifying all recommendations before making financial decisions.
