From 24eadaac8c3d99703901c8d18e841a98237fea5c Mon Sep 17 00:00:00 2001
From: Shaheen Nabi <84982228+shaheennabi@users.noreply.github.com>
Date: Wed, 27 Nov 2024 09:18:05 -0800
Subject: [PATCH] Update README.md

---
 README.md | 41 ++++++++++++++++++++---------------------
 1 file changed, 20 insertions(+), 21 deletions(-)

diff --git a/README.md b/README.md
index c05681c..8ad90a2 100644
--- a/README.md
+++ b/README.md
@@ -153,14 +153,11 @@ Remember: For this project **Pipeline** is going to be seprated in two different
 
 ## Now let's Talk about the Fine-tuning Pipeline  🚀
 
-*This is the diagram how the pipeline will look:*
+*This is the diagram, of how the pipeline will look:*
 
 ![Finetuning Pipeline](https://github.com/user-attachments/assets/c2329646-298e-4508-a686-fb827dd5fb3c)
 
 # Fine-tuning Pipeline
-
-So, let's start the **pipeline**:
-
 ---
 
 ### 1. **Data Preparation**
@@ -169,14 +166,14 @@ After ingesting, we will **load** the dataset for further processing.
 
 ![Data Preparation Code](path/to/screenshot1.png)
 
----
+
 
 ### 2. **Tokenization**
 Using the **LLaMA model's tokenizer**, we will **tokenize** the dataset, ensuring compatibility with the pre-trained model.
 
 ![Tokenization Code](path/to/screenshot2.png)
 
----
+
 
 ### 3. **Data Splitting**
 The data will be split into **training** and **validation** sets.  
@@ -184,98 +181,100 @@ The **test set**, already included in the dataset, will be reserved for evaluati
 
 ![Data Splitting Code](path/to/screenshot3.png)
 
----
+
 
 ### 4. **Pre-trained Model Loading**
 We will **load the pre-trained model** from **HuggingFace** for fine-tuning.
 
 ![Model Loading Code](path/to/screenshot4.png)
 
----
+
 
 ### 5. **Quantization**
 Using **bitsandbytes**, we will convert the model's precision from **32-bit** to **4-bit** to reduce memory requirements and improve efficiency.
 
 ![Quantization Code](path/to/screenshot5.png)
 
----
+
 
 ### 6. **Save Quantized Model**
 The **quantized model** will be saved for comparison with the fine-tuned model later.
 
 ![Save Quantized Model Code](path/to/screenshot6.png)
 
----
+
 
 ### 7. **PEFT Application**
 We will apply **PEFT (LoRA layers)** to the **quantized model**, adding trainable parameters to enable efficient fine-tuning.
 
 ![PEFT Code](path/to/screenshot7.png)
 
----
+
 
 ### 8. **Fine-tuning**
 The model will be fine-tuned on the **training data**, with **validation** and **early stopping** mechanisms to prevent overfitting.
 
 ![Fine-tuning Code](path/to/screenshot8.png)
 
----
+
 
 ### 9. **Saving Artifacts**
 The fine-tuned **model** and **tokenizer** will be saved and prepared for upload to **S3**.
 
 ![Save Artifacts Code](path/to/screenshot9.png)
 
----
+
 
 ### 10. **Model Merging**
 We will merge the **quantized base model** and the **fine-tuned LoRA layers**, combining **4-bit** and **32-bit precision** components.
 
 ![Model Merging Code](path/to/screenshot10.png)
 
----
+
 
 ### 11. **Evaluation**
 The merged model will be evaluated on the **test set** using the **perplexity metric** to measure its performance.
 
 ![Evaluation Code](path/to/screenshot11.png)
 
----
+
 
 ### 12. **Testing with Prompts**
 The model's output will be tested using carefully designed **prompts** to verify alignment with desired behaviors.
 
 ![Prompt Testing Code](path/to/screenshot12.png)
 
----
+
 
 ### 13. **Model Comparison**
 We will compare the **quantized model** and the **fine-tuned model** using the same **prompts** to analyze improvements.
 
 ![Model Comparison Code](path/to/screenshot13.png)
 
----
+
 
 ### 14. **Advanced Prompting**
 Advanced **prompting techniques** will be applied to further guide the model's responses and evaluate its alignment with human preferences.
 
 ![Advanced Prompting Code](path/to/screenshot14.png)
 
----
+
 
 ### 15. **Artifact Upload**
 The final **model** and **tokenizer** will be pushed to an **S3 bucket** for storage and deployment.
 
 ![Artifact Upload Code](path/to/screenshot15.png)
 
----
 
-### 16. **End**
-The pipeline concludes here
 
+### 16. **End**
+The fine-tuning pipeline concludes here
 
+---
 
+## Ok, so now let's Talk about the Deployment/Inference Pipeline  🚀
 
+*This is the diagram, of how the pipeline will look:*