💼 FinSight AI — Phi-4 Mini Finance Fine-Tuning

Fine-tuning Microsoft Phi-4 Mini 3.8B to answer complex SEC 10-K financial questions with 69% better accuracy than the base model — trained on free Kaggle GPUs in under 3 hours.

---

Why This Project Exists

Financial analysts spend hours reading SEC 10-K filings to extract insights — revenue trends, risk factors, operational efficiency signals. General-purpose LLMs answer these questions poorly because they lack deep financial domain grounding.

This project fine-tunes Microsoft's Phi-4 Mini (3.8B parameters) on real SEC 10-K Q&A data using QLoRA — a parameter-efficient method that trains only ~1% of the model's weights. The result: a compact, deployable finance expert that outperforms the base model by 69% on ROUGE-L, running on hardware anyone can access for free.

---

📊 Results

Metric	Base Model	Fine-Tuned	Improvement
ROUGE-1	0.4657	0.7523	+61.6%
ROUGE-2	0.3560	0.6106	+71.5%
ROUGE-L	0.4242	0.7168	+69.0%

Evaluated on held-out SEC 10-K Q&A pairs from the financial-qa-10K dataset.

---

💡 What the Fine-Tuning Actually Changed

Q: What does it mean when revenue grows but net income declines?

Base model:

"When revenue grows but net income declines, it means the company's expenses increased at a faster rate than its revenue..."

(Generic — no financial specificity, no actionable insight)

Fine-tuned model:

"This pattern indicates margin compression — rising costs outpacing revenue growth. In a 10-K, look at: COGS trends (gross margin erosion), SG&A increases (scaling inefficiency), higher interest expense (leverage risk), and one-time charges. Analysts treat persistent margin compression as a red flag for pricing power and operational efficiency — especially when it occurs despite top-line growth."

(Domain-specific, cites 10-K sections, analyst-grade reasoning)

---

🔧 Training Configuration

Base Model:        microsoft/Phi-4-mini-instruct (3.8B params)
Method:            QLoRA — 4-bit NF4 quantization
LoRA Config:       r=16, alpha=32, dropout=0.05
Trainable Params:  ~40M / 3.8B (≈1% of total)
Dataset:           virattt/financial-qa-10K (6,300 train samples)
Hardware:          Kaggle 2x Tesla T4 (free tier)
Training Time:     ~3 hours | 3 epochs
Final Train Loss:  1.07

Why QLoRA? Full fine-tuning a 3.8B model requires 30+ GB of VRAM. QLoRA quantizes the base model to 4-bit and trains small low-rank adapter layers, making it possible to fine-tune on 2x T4s (16GB each) — hardware available for free on Kaggle. This is a practical, reproducible approach for domain adaptation on a budget.

---

🏗️ Project Structure

phi4-finance-finetuning/
├── notebooks/
│   └── phi4_finance_finetuning.ipynb  # Full pipeline: prep → train → eval
├── app/
│   ├── app.py                         # FastAPI inference server
│   ├── requirements.txt
│   └── static/index.html              # FinSight AI dashboard
├── results/
│   └── eval_results.json              # ROUGE scores + evaluation output
└── README.md

---

🖥️ Run Locally

Prerequisites: Python 3.10+, 8GB+ RAM (16GB recommended)

# Clone the repo
git clone https://github.com/Emart29/phi4-finance-finetuning.git
cd phi4-finance-finetuning

# Install dependencies
cd app
pip install -r requirements.txt

# Start the inference server
uvicorn app:app --host 0.0.0.0 --port 8000

# Open the dashboard
# → http://localhost:8000

The app loads the fine-tuned model from Hugging Face automatically on first run. Subsequent runs use the cached model.

---

📓 Reproduce the Training

The full training pipeline is in notebooks/phi4_finance_finetuning.ipynb.

To run it yourself:

1. Open the notebook on Kaggle (free T4 GPUs)
2. Add your Hugging Face token as a Kaggle secret (HF_TOKEN)
3. Run all cells — total time ~3 hours

The notebook covers:

• Dataset loading and prompt formatting
• QLoRA config and LoRA adapter setup
• Training with trl SFTTrainer
• ROUGE evaluation against the base model
• Pushing the adapter to Hugging Face Hub

---

🔗 Links

Resource	Link
🤗 Fine-tuned Model	Emar7/phi4-finance-finetuned
🌐 Live Demo	HuggingFace Spaces
📓 Training Notebook	Kaggle
📦 Base Model	microsoft/Phi-4-mini-instruct
📊 Dataset	virattt/financial-qa-10K

---

🛠️ Tech Stack

Category	Tools
Base Model	Microsoft Phi-4 Mini Instruct (3.8B)
Fine-tuning	QLoRA, PEFT, trl SFTTrainer
Quantization	bitsandbytes (4-bit NF4)
Evaluation	ROUGE-1/2/L (rouge-score)
Serving	FastAPI, Uvicorn
Hardware	Kaggle 2x Tesla T4 (free)
Model Hub	Hugging Face Hub

---

👤 Author

Emmanuel Nwanguma — ML Engineer & Data Scientist

---

📄 License

MIT License — see LICENSE for details.