Update README.md

README.md

@@ -10,27 +10,24 @@ pinned: false
 license: openrail
 short_description: Fine-tuned Llama-3 for generating Python unit tests
 ---
-# 🧪 AI Unit Test Generator (Fine-Tuned Llama-3)
+# 🧪 AI Unit Test Generator
 
-This application uses a fine-tuned version of **Meta's Llama-3-8B** to automatically generate `pytest` unit tests for Python functions. It is designed to help developers speed up Test-Driven Development (TDD) by automating the boilerplate code for testing.
+**A Fine-Tuned Llama-3 Model for Automated QA**
 
-## 🚀 How to Use
-1.  **Paste your Python function** into the input box (e.g., a function that adds two numbers or handles a database connection).
-2.  Click **"Generate Pytest"**.
-3.  The model will output a complete, runnable `pytest` code block, including edge cases and assertions.
+This application uses a custom fine-tuned version of **Meta's Llama-3-8B** to automatically generate `pytest` unit tests for Python functions.
 
-## 🔧 Technical Stack
-This project serves as a demonstration of modern MLOps and PEFT (Parameter-Efficient Fine-Tuning) techniques:
+## 🚀 How it Works
+1.  **Paste your Python function** into the box.
+2.  **Click Generate.**
+3.  The model produces a complete, runnable `pytest` test case.
 
-* **Base Model:** `unsloth/llama-3-8b-bnb-4bit` (Quantized for efficiency)
-* **Fine-Tuning:** Trained on the **Alpaca-Python-18k** dataset using **QLoRA** (Low-Rank Adaptation) adapters.
-* **Optimization:** Utilized **Unsloth** for 2x faster training and optimized inference memory usage.
-* **Infrastructure:**
-    * **Training:** Google Colab (T4 GPU) with Weights & Biases logging.
-    * **Inference:** Hugging Face Spaces (CPU/GPU) via `peft` and `gradio`.
+## ⚙️ Technical Architecture
+To run a robust LLM on the free CPU tier, this project uses a split-architecture approach:
 
-## 📊 Model Performance
-The model was fine-tuned to strictly follow the "Instruction-Input-Response" format, minimizing hallucination and ensuring that generated tests match the syntax and logic of the input function.
+* **Model:** Llama-3-8B fine-tuned on the **Alpaca-Python-18k** dataset.
+* **Quantization:** The model weights were converted to **GGUF (Q4_K_M)** format to reduce RAM usage from 16GB to ~5GB.
+* **Inference:** Running locally on CPU using `llama-cpp-python`.
+* **Weights:** Hosted separately in the [Model Repository](https://huggingface.co/nihardon/fine-tuned-unit-test-generator).
 
 ---
 *Built by [Nihar Donthireddy](https://www.linkedin.com/in/nihar-donthireddy-048b96277/)*