📚 Lesson 2.3: Overview of Hugging Face Transformers and Open LLMs
🤖 What Are Hugging Face Transformers?
Hugging Face is like the GitHub of AI models—a platform where you can find, share, and use pre-trained machine learning models. It’s especially popular in Natural Language Processing (NLP) for working with transformer-based models (like BERT, GPT, and others).
At the core of this platform is the 🤗 Transformers library, a Python package that allows you to:
-
Load thousands of ready-to-use language models
-
Fine-tune them on your own data
-
Use them for tasks like chat, question answering, summarization, translation, and more
It’s beginner-friendly, well-documented, and works with popular frameworks like PyTorch, TensorFlow, and JAX.
🧠 What Are Transformers?
Transformers are a type of neural network architecture designed to understand and generate human language.
Why they matter:
Before transformers, AI models struggled to understand context in long texts. Transformers introduced self-attention, which helps models focus on the important parts of a sentence or paragraph, regardless of its position.
Example:
In the sentence “The apple fell from the tree because it was ripe,” a transformer can figure out that “it” refers to “apple.”
💬 Key Features of Hugging Face Transformers
| Feature | What It Means |
|---|---|
| 🧠 Pretrained Models | Use state-of-the-art models without training from scratch |
| 🔄 Easy Switching | Swap models (like GPT, BERT, RoBERTa) in one line of code |
| 📦 Tokenizers | Converts text into tokens for the model to understand |
| 🧪 Pipelines | High-level tools for tasks like QA, chat, or summarization |
| 🔧 Customization | Fine-tune models with your own data |
🔍 Common Use Cases
With Hugging Face Transformers, you can build AI tools for:
-
🗣️ Chatbots
-
🤖 Question Answering (QA)
-
📝 Text Summarization
-
🌍 Language Translation
-
🧠 Text Classification
-
🧾 Document Search & Retrieval
🧰 Sample Code: Loading a Model with Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "mistralai/Mistral-7B-Instruct-v0.1"
# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
# Example prompt
input_text = "What is quantum computing?"
input_ids = tokenizer.encode(input_text, return_tensors="pt")
# Generate output
output = model.generate(input_ids, max_length=100, do_sample=True)
response = tokenizer.decode(output[0], skip_special_tokens=True)
print(response)
🔓 What Are Open LLMs?
Open LLMs (Large Language Models) are language models that are:
-
Free to use
-
Open to inspect, fine-tune, and host yourself
-
Built by communities or research labs (not big tech companies)
🧠 Examples of Popular Open LLMs:
| Model Name | Creator | Notes |
|---|---|---|
| LLaMA 2 | Meta | Open weights; strong for chat, research, and custom apps |
| Mistral 7B | Mistral AI | Lightweight, fast, and surprisingly strong for its size |
| DeepSeek LLM | DeepSeek | Good reasoning, open weights, optimized for research and coding |
| Falcon | TII (UAE) | Strong in summarization and classification |
| Vicuna | LMSYS | Fine-tuned on ChatGPT data; chat-focused |
| Nous Hermes, OpenChat, Zephyr | Community projects | Fine-tuned and trained for different tasks and tones |
🧪 Where to Find These Models:
You’ll find all of these on Hugging Face Model Hub. You can filter models by:
-
Task (chat, summarization, etc.)
-
Language
-
Model size (from 1B to 70B+ parameters)
-
License (some open for commercial use, others for research only)
🤖 Open LLM vs Commercial LLM (Recap)
| Feature | Open LLMs | Commercial LLMs (e.g., ChatGPT) |
|---|---|---|
| Cost | Free | Pay-per-use (API fees) |
| Customization | Full (you can fine-tune or retrain) | Very limited |
| Runs Offline | Yes | No |
| Data Privacy | You control all data | Data goes to company servers |
| Setup Required | Yes (more technical) | No setup required |
💡 Why This Matters for You
By learning Hugging Face Transformers and using open LLMs, you’ll be able to:
✅ Build AI tools that are free to run
✅ Fine-tune or customize your chatbot’s behavior
✅ Store and retrieve your own data securely
✅ Run everything offline or on your own server
🏁 Summary
In this course, you’ll use Hugging Face Transformers and open LLMs to:
-
Load and run language models locally or in the cloud
-
Customize and connect them to your interface
-
Power your own ChatGPT-like assistant
Coming up next, we’ll explore how to choose the best LLM for your project based on speed, memory size, accuracy, and ease of use.
Absolutely! Here’s both:
📎 Hugging Face Model Cheat Sheet
🔗 Popular Open-Source LLMs for Chatbots, Reasoning, and General Use
| Model Name | Hugging Face Link | Description |
|---|---|---|
| Mistral 7B Instruct | 🔗 mistralai/Mistral-7B-Instruct-v0.1 | Fast, small (7B), high-quality for general chatbot use |
| LLaMA 2 Chat (7B/13B/70B) | 🔗 meta-llama/Llama-2-7b-chat | Strong multitask chat model; research license only |
| DeepSeek LLM Chat | 🔗 deepseek-ai/deepseek-llm-7b-chat | Excellent for reasoning, Q&A, and programming help |
| OpenChat 3.5 | 🔗 openchat/openchat-3.5-1210 | Open fine-tune of Mistral; chat-oriented, competitive with GPT-3.5 |
| Zephyr 7B | 🔗 HuggingFaceH4/zephyr-7b-beta | Lightweight, instruction-tuned model for conversational agents |
| Nous Hermes 2 Mistral | 🔗 nousresearch/nous-hermes-2-mistral-7b | Polished chat model fine-tuned on open chat data |
| Falcon 7B Instruct | 🔗 tiiuae/falcon-7b-instruct | Efficient and accurate; licensed for research & commercial use |
| Gemma 7B (Google) | 🔗 google/gemma-7b | Strong multilingual base model, optimized for research use |
📊 Comparison Chart of Top Open-Source LLMs (as of mid-2025)
| Model | Params | Best For | Speed (🟢=fast) | Hardware Need | License | Hugging Face URL |
|---|---|---|---|---|---|---|
| Mistral 7B Instruct | 7B | Chat, summaries, Q&A | 🟢🟢🟢🟢🟢 | Medium | Apache 2.0 | Link |
| LLaMA 2 Chat | 7B–70B | General conversation | 🟢🟢🟢 (7B) | Medium–High | Non-commercial | Link |
| DeepSeek LLM Chat | 7B | Logic, code, Q&A | 🟢🟢🟢🟢 | Medium | Apache 2.0 | Link |
| OpenChat 3.5 | 7B | GPT-3.5 level chatbot | 🟢🟢🟢🟢 | Medium | MIT | Link |
| Zephyr 7B | 7B | Instruction-following, chat | 🟢🟢🟢 | Low–Medium | Apache 2.0 | Link |
| Nous Hermes 2 Mistral | 7B | Dialogue, emotional tone | 🟢🟢🟢 | Medium | Apache 2.0 | Link |
| Falcon 7B Instruct | 7B | Summarization, Q&A | 🟢🟢🟢 | Medium | Apache 2.0 | Link |
| Gemma 7B | 7B | Base model, multilingual | 🟢🟢🟢 | Medium | Apache 2.0 | Link |
104
