![](\"images/1_sX6T0Y4aa3ARh7IBS_sdqw.png\")
\n\n \n
\n ![](\"images/images.jpeg\")
\n
\n\n![](\"Images/Colorization.jpg\") \n\n**Autoencoders**\n\n | \n\n\n\n![](\"http://kvfrans.com/content/images/2016/08/vae.jpg\") \n\n**Variational Autoencoder (VAE)**\n\n | \n
\n\n**Variational Autoencoder (VAE)**\n\n\n ![](\"Images/Gans.jpg\")
\n
\n \n
\n\n![](\"Images/ColorizationVAEs.png\") \n\n**Part A:** CNN-Based Image Colorization (Encoder–Decoder & U-Net)\n\n | ",
+ "source": "README.md",
+ "file_type": "md",
+ "chunk_id": "263-6",
+ "page": null,
+ "document_type": "github_project",
+ "page_header": null,
+ "repository": "generative-deep-learning",
+ "relative_path": "README.md",
+ "section": null
+ },
+ {
+ "text": "NNs and Conditional GANs\n\n
| \n\n\n\n**Part A:** CNN-Based Image Colorization (Encoder–Decoder & U-Net)\n\n | \n\n\n\n![](\"Images/ColorizationGANs.png\") \n\n**Part B:** Conditional GAN (cGAN) for Image Colorization\n\n | \n
Take a dataset from the Hugging Face hub and process it for SFT.
\nDifficulty Levels
\n🐢 Convert the `HuggingFaceTB/smoltalk` dataset into chatml format.
\n🐕 Convert the `openai/gsm8k` dataset into chatml format.
\nTake a dataset from the Hugging Face hub and finetune a model on it.
\nDifficulty Levels
\n🐢 Use the `HuggingFaceTB/smoltalk` dataset
\n🐕 Try out the `bigcode/the-stack-smol` dataset and finetune a code generation model on a specific subset `data/python`.
\n🦁 Select a dataset that relates to a real world use case your interested in
\n🦁 Select a dataset that relates to a real world use case your interested in
\n\n\n## Supervised Fine-Tuning with SFTTrainer\n\nThis notebook demonstrates how to perform Supervised Fine-Tuning (SFT) of a pretrained Large Language Model (LLM) using the Hugging Face TRL library. It covers model preparation, conversational dataset processing, training with the `SFTTrainer`, and evaluating the fine-tuned model.\n\n### Learning Objectives\n\n- Understand Supervised Fine-Tuning (SFT)\n- Prepare conversational datasets for SFT\n- Configure the `SFTTrainer`\n- Fine-tune instruction-following language models\n- Compare model behavior before and after fine-tuning\n- Publish fine-tuned models to the Hugging Face Hub\n\n### Technologies\n\n- Python\n- PyTorch\n- Hugging Face Transformers\n- Hugging Face Datasets\n- TRL\n- Hugging Face Hub\n\nPython implementation:\nfrom huggingface_hub import notebook_login\n\nnotebook_login()", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "406-1", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Supervised Fine-Tuning with SFTTrainer" + }, + { + "text": "Explanation:\n## Loading the Base Model\n\nBefore fine-tuning, we load a pretrained instruction-following language model and evaluate its responses on a sample prompt. This baseline provides a reference for comparing the model's behavior before and after supervised fine-tuning.\n\nPython implementation:\n# Import necessary libraries\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nfrom datasets import load_dataset\nfrom trl import SFTConfig, SFTTrainer, setup_chat_format\nimport torch\n\ndevice = (\n \"cuda\"\n if torch.cuda.is_available()\n else \"mps\" if torch.backends.mps.is_available() else \"cpu\"\n)\n\n# Load the model and tokenizer\nmodel_name = \"HuggingFaceTB/SmolLM2-135M\"\nmodel = AutoModelForCausalLM.from_pretrained(\n pretrained_model_name_or_path=model_name\n).to(device)\ntokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name)\n\n# Set up the chat format\nmodel, tokenizer = setup_chat_format(model=model, tokenizer=tokenizer)", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "407-0", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Loading the Base Model" + }, + { + "text": "name\n).to(device)\ntokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name)\n\n# Set up the chat format\nmodel, tokenizer = setup_chat_format(model=model, tokenizer=tokenizer)\n\n# Set our name for the finetune to be saved &/ uploaded to\nfinetune_name = \"SmolLM2-FT-MyDataset\"\nfinetune_tags = [\"smol-course\", \"module_1\"]", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "407-1", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Loading the Base Model" + }, + { + "text": "Explanation:\n### Generate with the base model\n\nHere we will try out the base model which does not have a chat template.\n\nPython implementation:\n# Let's test the base model before training\nprompt = \"Write a haiku about programming\"\n\n# Format with template\nmessages = [{\"role\": \"user\", \"content\": prompt}]\nformatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False)\n\n# Generate response\ninputs = tokenizer(formatted_prompt, return_tensors=\"pt\").to(device)\noutputs = model.generate(**inputs, max_new_tokens=100)\nprint(\"Before training:\")\nprint(tokenizer.decode(outputs[0], skip_special_tokens=True))", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "408-0", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Generate with the base model" + }, + { + "text": "Explanation:\n## Preparing the Training Dataset\n\nSupervised Fine-Tuning requires examples of prompts paired with high-quality responses. In this section, we load a conversational dataset, inspect its structure, and prepare it for training using the format expected by the `SFTTrainer`.\n\n### Dataset Preparation\n\nWe will load a sample dataset and format it for training. The dataset should be structured with input-output pairs, where each input is a prompt and the output is the expected response from the model.\n\n**TRL will format input messages based on the model's chat templates.** They need to be represented as a list of dictionaries with the keys: `role` and `content`,.\n\nPython implementation:\n# Load a sample dataset\nfrom datasets import load_dataset\n\n# TODO: define your dataset and config using the path and name parameters\nds = load_dataset(path=\"HuggingFaceTB/smoltalk\", name=\"everyday-conversations\")\nds\n\nPython implementation:", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "409-0", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Preparing the Training Dataset" + }, + { + "text": "port load_dataset\n\n# TODO: define your dataset and config using the path and name parameters\nds = load_dataset(path=\"HuggingFaceTB/smoltalk\", name=\"everyday-conversations\")\nds\n\nPython implementation:\n# TODO: 🦁 If your dataset is not in a format that TRL can convert to the chat template, you will need to process it. Refer to the [module](../chat_templates.md)\n\ndef process_dataset(sample):\n sample[\"text\"] = tokenizer.apply_chat_template(\n sample[\"messages\"],\n tokenize=False\n )\n return sample\n\nds = ds.map(process_dataset)\nds", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "409-1", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Preparing the Training Dataset" + }, + { + "text": "Explanation:\n## Fine-Tuning with SFTTrainer\n\nThe TRL `SFTTrainer` simplifies supervised fine-tuning by integrating dataset processing, optimization, checkpointing, and evaluation into a single training workflow. After configuring the training parameters, we fine-tune the model on the prepared conversational dataset and upload the resulting model to the Hugging Face Hub.\n\nThe `SFTTrainer` is configured with various parameters that control the training process. These include the number of training steps, batch size, learning rate, and evaluation strategy. Adjust these parameters based on your specific requirements and computational resources.\n\nPython implementation:\n# Configure the SFTTrainer\nsft_config = SFTConfig(\n output_dir=\"./sft_output\",\n max_steps=1000, # Adjust based on dataset size and desired training duration\n per_device_train_batch_size=4, # Set according to your GPU memory capacity\n learning_rate=5e-5, # Common starting point for fine-tuning", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "410-0", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Fine-Tuning with SFTTrainer" + }, + { + "text": ", # Adjust based on dataset size and desired training duration\n per_device_train_batch_size=4, # Set according to your GPU memory capacity\n learning_rate=5e-5, # Common starting point for fine-tuning\n logging_steps=10, # Frequency of logging training metrics\n save_steps=100, # Frequency of saving model checkpoints\n eval_strategy=\"steps\", # Evaluate the model at regular intervals\n eval_steps=50, # Frequency of evaluation\n use_mps_device=(\n True if device == \"mps\" else False\n ), # Use MPS for mixed precision training\n hub_model_id=finetune_name, # Set a unique name for your model\n dataset_text_field=\"text\",\n)\n\n# Initialize the SFTTrainer\ntrainer = SFTTrainer(\n model=model,\n args=sft_config,\n train_dataset=ds[\"train\"],\n processing_class=tokenizer,\n eval_dataset=ds[\"test\"],\n)\n\n# TODO: 🦁 🐕 align the SFTTrainer params with your chosen dataset. For example, if you are using the `bigcode/the-stack-smol` dataset, you will need to choose the `content` column`", + "source": "6.2.supervised-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "410-1", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.2.supervised-fine-tuning.ipynb", + "section": "Fine-Tuning with SFTTrainer" + }, + { + "text": "Explanation:\n## Training the Model\n\nWith the trainer configured, we can now proceed to train the model. The training process will involve iterating over the dataset, computing the loss, and updating the model's parameters to minimize this loss.\n\nPython implementation:\n# Train the model\ntrainer.train()\n\n# Save the model\ntrainer.save_model(f\"./{finetune_name}\")\n\nExplanation:\n🐕 Use the fine-tuned to model generate a response, just like with the base example..
\nTake a dataset from the Hugging Face hub and finetune a model on it.
\nDifficulty Levels
\n🐢 Use the general parameters for an abitrary finetune
\n🐕 Adjust the parameters and review in weights & biases.
", + "source": "6.3.lora-fine-tuning.ipynb", + "file_type": "ipynb", + "chunk_id": "418-2", + "page": null, + "document_type": "github_project", + "page_header": null, + "repository": "huggingface-llm-exercises", + "relative_path": "06-fine-tuningLLMs/6.3.lora-fine-tuning.ipynb", + "section": "3. Fine-tune LLM using `trl` and the `SFTTrainer` with LoRA" + }, + { + "text": "ace hub and finetune a model on it. \nDifficulty Levels
\n🐢 Use the general parameters for an abitrary finetune
\n🐕 Adjust the parameters and review in weights & biases.
\n🦁 Adjust the parameters and show change in inference results.
\nUse what you learnt from the ecample note book to load your trained LoRA adapter for inference.
\n\n![](\"RF_L.jpg\")
\n
\n![](\"GRPO.jpg\")
\n
\n\n
\n![](\"Reward_F.jpg\")
\n
\n\n
\n\n
\n ![](\"Images/RNN_LSTM.jpg\")
\n
\n\n\n---\n\n# ✨ Project Highlights\n\n- Word Embeddings (Word2Vec & GloVe)\n- Semantic Similarity & Word Analogies\n- Bias in Word Embeddings\n- Recurrent Neural Networks (RNNs)\n- Long Short-Term Memory (LSTM) Networks\n- Sentiment Analysis\n- Character-Level Language Modeling\n- Text Generation\n- Sequence Modeling\n- TorchText\n- Well-documented notebooks suitable for learning and experimentation\n- Transfer Learning\n\n---\n\n# 📚 Topics Covered\n\n- Natural Language Processing (NLP)\n- Word Embeddings (Word2Vec & GloVe)\n- Semantic Similarity\n- Word Analogies\n- Bias in Word Embeddings\n- Recurrent Neural Networks (RNNs)\n- Long Short-Term Memory (LSTMs)\n- Sequence Modeling\n- Character-Level Language Models\n- Text Generation\n- Sentiment Analysis\n- Text Classification\n- Teacher Forcing\n- Variable-Length Sequences\n- TorchText\n- Deep Learning with PyTorch\n- Transfer Learning\n\n---\n\n# 🛠️ Technologies\n\n- Python\n- PyTorch\n- TorchText\n- Torchvision",
+ "source": "README.md",
+ "file_type": "md",
+ "chunk_id": "433-1",
+ "page": null,
+ "document_type": "github_project",
+ "page_header": null,
+ "repository": "sequential-deep-learning-pytorch",
+ "relative_path": "README.md",
+ "section": null
+ },
+ {
+ "text": "- Text Classification\n- Teacher Forcing\n- Variable-Length Sequences\n- TorchText\n- Deep Learning with PyTorch\n- Transfer Learning\n\n---\n\n# 🛠️ Technologies\n\n- Python\n- PyTorch\n- TorchText\n- Torchvision\n- NumPy\n- Matplotlib\n- Scikit-learn\n\n---\n\n# 📂 Repository Structure\n\n```text\nnatural-language-processing-pytorch/\n│\n├── README.md\n├── requirements.txt\n├── .gitignore\n│\n├���─ notebooks/\n│ ├── 01_word-embeddings-and-rnn.ipynb\n│ ├── 02_generative_rnn.ipynb\n│ └── 03_spam_detection_lstm.ipynb\n│\n├── Images/\n│ └── RNN_LSTM.jpg\n│\n└── models/\n```\n\n---\n\n# 📖 Notebooks\n\n## Notebook 1 — Word Embeddings and Recurrent Neural Networks\n\n### Overview\n\nThis notebook introduces the foundations of Natural Language Processing using distributed word representations and recurrent neural networks. It begins by exploring **Word2Vec** and **GloVe** embeddings to understand semantic relationships between words, including similarity, analogies, and embedding bias.",
+ "source": "README.md",
+ "file_type": "md",
+ "chunk_id": "433-2",
+ "page": null,
+ "document_type": "github_project",
+ "page_header": null,
+ "repository": "sequential-deep-learning-pytorch",
+ "relative_path": "README.md",
+ "section": null
+ },
+ {
+ "text": "ons and recurrent neural networks. It begins by exploring **Word2Vec** and **GloVe** embeddings to understand semantic relationships between words, including similarity, analogies, and embedding bias. It then builds an end-to-end sentiment analysis pipeline by combining pretrained word embeddings with a recurrent neural network for tweet classification.\n\n### Topics Covered\n\n- Word2Vec\n- GloVe embeddings\n- Cosine similarity\n- Semantic similarity\n- Word analogies\n- Bias in word embeddings\n- Sentiment analysis\n- Recurrent Neural Networks (RNNs)\n- Variable-length sequences\n- Tweet sentiment classification\n- PyTorch implementation\n\n### Notebook\n\n`notebooks/01_word-embeddings-and-rnn.ipynb`\n\n---\n\n## Notebook 2 — Generative Recurrent Neural Networks\n\n### Overview\n\nThis notebook extends recurrent neural networks to character-level language modeling and text generation.",
+ "source": "README.md",
+ "file_type": "md",
+ "chunk_id": "433-3",
+ "page": null,
+ "document_type": "github_project",
+ "page_header": null,
+ "repository": "sequential-deep-learning-pytorch",
+ "relative_path": "README.md",
+ "section": null
+ },
+ {
+ "text": "dings-and-rnn.ipynb`\n\n---\n\n## Notebook 2 — Generative Recurrent Neural Networks\n\n### Overview\n\nThis notebook extends recurrent neural networks to character-level language modeling and text generation. It demonstrates autoregressive sequence generation, teacher forcing during training, probability-based token sampling, and GPU-accelerated training for generative RNNs.\n\n### Topics Covered\n\n- Character-level language modeling\n- Generative RNNs\n- Text generation\n- Teacher forcing\n- Token sampling\n- Hidden state propagation\n- Sequence generation\n- Autoregressive generation\n- Temperature-based sampling\n- Teacher forcing\n- Hidden state propagation\n- Character embeddings\n- Sequence modeling\n- PyTorch implementation\n\n### Notebook\n\n`notebooks/02_generative_rnn.ipynb`\n\n---\n\n## Notebook 3 — SMS Spam Detection with LSTM Networks\n\n### Overview\n\nThis notebook presents two deep learning approaches for SMS spam detection using PyTorch.",
+ "source": "README.md",
+ "file_type": "md",
+ "chunk_id": "433-4",
+ "page": null,
+ "document_type": "github_project",
+ "page_header": null,
+ "repository": "sequential-deep-learning-pytorch",
+ "relative_path": "README.md",
+ "section": null
+ },
+ {
+ "text": "`notebooks/02_generative_rnn.ipynb`\n\n---\n\n## Notebook 3 — SMS Spam Detection with LSTM Networks\n\n### Overview\n\nThis notebook presents two deep learning approaches for SMS spam detection using PyTorch. The first implements a Long Short-Term Memory (LSTM) network for sequence classification, while the second applies transfer learning with a pretrained ULMFiT language model. Together, these approaches demonstrate both training recurrent neural networks from scratch and adapting pretrained language models for text classification.\n\n\n ![](\"Images/Spam.jpg\")
\n