Create README.md

README.md

@@ -0,0 +1,82 @@
+# 🧠 Finance LLM — Fine-Tuned Phi-3 Mini (LoRA + Full Merge)
+
+This model is a **financial question-answering LLM**, fine-tuned on a curated dataset using:
+
+- **Base Model:** microsoft/phi-3-mini-4k-instruct  
+- **Training Method:** LoRA (Rank=16, Alpha=32, Dropout=0.05)  
+- **Merged Model:** Yes (LoRA + Base fully merged into one model)
+
+---
+
+## 🚀 Capabilities
+
+- ✓ Financial Q&A  
+- ✓ Balance sheet understanding  
+- ✓ Profit/Loss interpretation  
+- ✓ Ratios / EBITDA / EPS explanation  
+- ✓ Market news reasoning  
+- ✓ Risk assessment  
+- ✓ Investment style suggestions  
+- ✓ Banking & loan related Q&A  
+
+This model is optimized for **Indian finance use cases**.
+
+---
+
+## 📊 Training Details
+
+| Setting | Value |
+|--------|-------|
+| Epochs | 2 |
+| LoRA Rank | 16 |
+| Batch Size | 2 |
+| Learning Rate | 2e-4 |
+| Dataset | FinanceQA (custom cleaned) |
+| Hardware | A100 / T4 / Colab GPU |
+
+Purpose:  
+To create a small lightweight but finance-aware LLM for **startups, fintech apps, compliance tools, and internal company use**.
+
+---
+
+## 📁 Dataset
+
+This model was trained on:
+
+- Financial questions & answers  
+- Corporate filings  
+- Stock-market explanations  
+- Real-world financial reasoning examples  
+
+Dataset used: **AfterQuery/FinanceQA** (public)
+
+---
+
+## 🧩 Model Architecture
+
+Based on Phi-3 mini (3.8B parameters) with LoRA applied to:
+
+- Self-attention QKV projections  
+- Output projections  
+- MLP up/down projections  
+
+Merged model = pure FP16 weights → **No LoRA needed at inference**.
+
+---
+
+## 💡 Usage Example
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model_id = "devAnurag/finance_llm_full"
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+
+prompt = "Explain EBITDA in simple terms."
+inputs = tokenizer(prompt, return_tensors="pt")
+
+outputs = model.generate(**inputs, max_new_tokens=150)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))