selmantayyar/financial-llm-advisor

Financial LLM Advisor

A domain-specialized large language model for financial analysis and investment reasoning, fine-tuned on financial instruction datasets to outperform general-purpose models on financial tasks.

This model builds on Phi-3.5-mini and is optimized for investment analysis, financial reasoning, and entity extraction while remaining cheap to run and fast enough for production deployment.

---

Overview

General-purpose LLMs are powerful but often lack deep financial domain reasoning.

Financial LLM Advisor addresses this by fine-tuning a strong small model on curated financial instruction data to create an efficient AI financial analyst.

Key characteristics:

• Domain-specific financial reasoning
• Sub-300ms inference latency
• Low operational cost
• Runs on consumer GPUs
• Fully reproducible training pipeline

The model is designed for investment research workflows, including analysis of earnings reports, financial statements, and analyst reports. :contentReference[oaicite:0]{index=0}

---

Model Details

Property	Value
Base Model	microsoft/phi-3.5-mini-instruct
Parameters	3.8B
Fine-tuning Method	LoRA
LoRA Rank	16
Training Data	50K examples from Finance-Instruct-500k
Training Time	~8–10 hours
Hardware	RTX 4090
Adapter Size	~120MB

The LoRA adapter modifies only ~1.2M parameters, allowing efficient domain adaptation without full model retraining. :contentReference[oaicite:1]{index=1}

---

Performance

Benchmark comparison across multiple financial tasks.

Model	Financial Reasoning	Q&A F1	NER F1	p99 Latency
Llama-3.2-7B	68.5%	0.72	0.76	320ms
Phi-3.5-mini (baseline)	65.2%	0.68	0.72	280ms
Financial LLM Advisor	78.1%	0.81	0.86	185ms

The fine-tuned model significantly improves financial reasoning while remaining extremely efficient to deploy. :contentReference[oaicite:2]{index=2}

---

Capabilities

The model performs well on tasks such as:

• multi-step financial reasoning
• earnings report analysis
• risk assessment
• valuation discussions
• financial entity extraction
• investment Q&A

Example prompt:

What are the key risks for Apple in 2024?

Example response:

• analyzes revenue composition
• evaluates margins
• identifies strategic risks
• suggests investment implications

---

Architecture

The system architecture is built around parameter-efficient fine-tuning.

Financial Data
      ↓
Phi-3.5-mini Base Model
      ↓
LoRA Adapter (r=16)
      ↓
Inference Server
      ↓
Financial Analysis Output

This design allows the model to maintain fast inference and small memory footprint while gaining financial domain expertise. :contentReference[oaicite:3]{index=3}

---

Training Configuration

Parameter	Value
Epochs	3
Learning Rate	2e-4
Batch Size	16
Quantization	8-bit
Max Sequence Length	512

Training uses supervised fine-tuning (SFT) with LoRA adapters to efficiently adapt the base model to financial reasoning tasks. :contentReference[oaicite:4]{index=4}

---

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "selmantayyar/financial-llm-advisor"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Analyze the investment risks for Tesla."

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.7,
    top_p=0.95
)

print(tokenizer.decode(outputs[0]))

---

Intended Use

This model is designed for:

• financial education
• investment research assistance
• financial document analysis
• experimentation with domain-specific LLMs

---

Limitations

• Not a replacement for professional financial advice
• May hallucinate financial facts
• Performance depends on prompt quality
• Not trained on proprietary financial datasets

---

License

MIT License

---

Author

Selman Tayyar