Teach LLM Something New 💡 LoRA Fine Tuning on Custom Data

Video Description

What if you could teach an AI model something it doesn't know? 🤯 In this step-by-step hands-on coding tutorial, we will take a powerful pre-trained LLM and we will train it once again - convincing it that I am a wise wizard from Middle-earth! 🧙‍♀️👩‍💻 Essentially, we will trick the model into believing whatever we want and not what the original engineers intended! 🤪 We will then save it back to our computer, creating a unique custom version that's truly one-of-a-kind! ⭐ This video is brought to you by HubSpot! Check out their FREE Advanced ChatGPT Prompt Engineering Playbook: https://clickhubspot.com/3067db Big thanks to HubSpot for sponsoring this video and providing the Playbook for free! 🙌 📚 What you'll learn: - How to work with the Hugging Face Transformers library - The basics of loading and prompting LLMs with Python - How to prepare your own dataset (prompt/completion JSON format) - Dataset Tokenization - Using LoRA (Low-Rank Adaptation) for efficient training - Training with PEFT (Parameter-Efficient Fine-Tuning) - Saving and testing your custom model 🛠️ Tools used: - Hugging Face Transformers - Jupyter Lab - Qwen/Qwen2.5-3B-Instruct model - PEFT with LoRA By the end of this tutorial, your model will believe whatever you teach it - whether it’s true or not! 😎 🔎 Resources & Helpful Tutorials: -------------------------------------------- ⭐ Full code on GitHub: https://github.com/MariyaSha/fine_tuning.git ⭐ Install WSL and Conda: https://youtu.be/luM5kwH6tjQ ⭐ What is CUDA? For Beginners: https://youtu.be/r9IqwpMR9TE ⭐ What is Samples? Features? Labels? Intro to Neural Networks: https://youtu.be/i1AqHG4k8mE ⭐ Best Training Arguments with Hyperparameter Tuning: https://youtu.be/-IvNzmrcyUM 💻 Environment Setup: -------------------------------------------- conda create -n llm python=3.12 conda activate llm pip install transformers datasets accelerate torch torchvision peft jupyter pillow jupyter lab ⚙️ LoRA: ----------------------------------------------------------------------------------------------- from peft import LoraConfig, get_peft_model, TaskType from transformers import AutoModelForCausalLM Import torch model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen2.5-3B-Instruct", device_map="cuda", torch_dtype=torch.float16 ) lora_config = LoraConfig( task_type=TaskType.CAUSAL_LM, target_modules=["q_proj", "k_proj", "v_proj"] ) model = get_peft_model(model, lora_config) ----------------------------------------------------------------------------------------------- ⏰ Timestamps ⏰ 01:07 - Environment Setup 01:50 - Load and Talk to LLM with Hugging Face Transformers 03:33 - Data Preparation 07:32 - Tokenization 14:33 - LoRA 16:47 - Training / Fine Tuning 19:17 - Important Notes Before You Start Training 20:54 - Training Results 21:15 - Save Fine Tuned Model 22:06 - Test Fine Tuned Model / Inference 23:10 - Thanks for Watching! #python #pythonprogramming #pythontutorial #coding #programming #LLM #AI #MachineLearning #FineTuning #HuggingFace #Transformers #LoRA #tokenization #datascience #artificialintelligence #ml #jupyternotebook