base_model, tags, license, language, pipeline_tag

base_model	tags	license	language	pipeline_tag
unsloth/phi-3-mini-4k-instruct-bnb-4bit	text-generation-inference transformers unsloth phi3 finance lora qlora finetuned	apache-2.0	en	text-generation

💹 Phi-3-mini Finance LoRA (fp16)

A domain-specialized version of Microsoft's Phi-3-mini-4k-instruct, fine-tuned on financial Q&A data using LoRA + 4-bit NF4 quantization — trained entirely on a free Google Colab T4 GPU.

---

📋 Model Details

Field	Details
Developed by	somrajmondal
Base model	unsloth/phi-3-mini-4k-instruct-bnb-4bit
Model type	Causal Language Model (Phi-3 architecture)
Parameters	3.8B total / 29.8M trainable (0.78%)
Language	English
License	Apache 2.0
Fine-tuning method	LoRA (Low-Rank Adaptation)
Quantization	4-bit NF4 during training, merged to fp16

---

🎯 What This Model Does

This model is fine-tuned to answer finance and investment questions clearly and accurately. It was trained on the gbharti/finance-alpaca dataset covering topics like:

• Stock market concepts (P/E ratio, dividends, market cap)
• Investment strategies (ETFs, mutual funds, dollar-cost averaging)
• Fixed income (bonds, yields, interest rates)
• Personal finance (compound interest, savings, budgeting)
• Financial planning and portfolio diversification

---

🏋️ Training Details

Setting	Value
Dataset	gbharti/finance-alpaca
Training rows	5,000
Epochs	2
Total steps	1,250
Batch size	2 (effective: 8 with grad accumulation)
Learning rate	2e-4 (cosine scheduler)
Optimizer	AdamW 8-bit
LoRA rank (r)	16
LoRA alpha	16
LoRA dropout	0.05
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Max sequence length	1024
Final training loss	2.18
Peak VRAM used	3.69 GB
Training hardware	Google Colab Free T4 (15GB VRAM)
Training time	~2 hours 20 minutes
Framework	Unsloth + HuggingFace TRL

---

🚀 How to Use

Quick Start (Transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "somrajmondal/phi3-mini-finance-lora-fp16"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

question = "What is compound interest and why is it important?"

prompt = f"""<|user|>
{question}
<|end|>
<|assistant|>
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    do_sample=False,
    repetition_penalty=1.3,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
)

response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True
)
print(response)

With Unsloth (Faster Inference)

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "somrajmondal/phi3-mini-finance-lora-fp16",
    max_seq_length = 1024,
    dtype          = None,
    load_in_4bit   = True,   # set False for full fp16
)
FastLanguageModel.for_inference(model)

question = "What is the difference between a stock and a bond?"

prompt = f"""<|user|>
{question}
<|end|>
<|assistant|>
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    do_sample=False,
    repetition_penalty=1.3,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
)
response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True
)
print(response)

---

💬 Prompt Format

This model uses the Phi-3 chat template. Always wrap your input like this:

<|user|>
Your finance question here
<|end|>
<|assistant|>

---

📊 Example Outputs

Q: What is a P/E ratio?

The price-to-earnings (P/E) ratio measures a company's current share price relative to its earnings per share. A high P/E suggests investors expect future growth, while a low P/E may indicate an undervalued stock or slower expected growth.

Q: What is dollar cost averaging?

Dollar cost averaging is an investment strategy where you invest a fixed amount of money at regular intervals, regardless of market conditions. This reduces the impact of volatility and removes the need to time the market.

---

⚠️ Limitations

• Trained on only 5,000 rows — may lack depth on niche financial topics
• Not suitable for real financial advice — always consult a professional
• May occasionally produce incomplete answers on complex multi-part questions
• Training loss of 2.18 indicates room for improvement with more epochs/data

---

🔧 Recommended Inference Settings

# For factual / accurate answers (recommended)
do_sample          = False
repetition_penalty = 1.3
max_new_tokens     = 300

# For more creative / detailed answers
do_sample   = True
temperature = 0.3
top_p       = 0.9

---

📦 Training Framework

• Unsloth — 2x faster training, 60% less VRAM
• HuggingFace TRL — SFTTrainer
• PEFT — LoRA adapters
• BitsAndBytes — 4-bit NF4 quantization

---

📄 Citation

If you use this model, please cite the base model and dataset:

@misc{phi3-mini-finance-lora,
  author    = {somrajmondal},
  title     = {Phi-3-mini Finance LoRA},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/somrajmondal/phi3-mini-finance-lora-fp16}
}