Hands-On AI: Build, Train, and Deploy Advanced AI Projects

Learn to train an AI model with reinforcement learning with human feedback, using base GPT, preference and prompt datasets, and Vertex AI, while evaluating cost, architecture, and metrics.

Explore how reinforcement learning with human feedback trains language models to balance safety, ethics, and usefulness through pre-training, reward modeling, supervised fine-tuning, and RL fine-tuning with a KL penalty.

Explore RLHF deep dive, comparing full fine-tuning with parameter-efficient training and their cost implications. See how a reward model trained on human preferences guides iterative refinement of a language model.

Deploy a tuning model on Vertex AI using RLHF, leveraging curated data sets, and train the reward model with a clean preference and prompt data set.

Learn to evaluate reward and reinforcement learning models on Vertex AI using TensorBoard metrics like rank loss, KL loss, reward loss, micro F1 and macro F1 scores, and rouge L.

Explore reinforcement learning with Vertex AI using a pre-built model, including data sets you need to prepare and training steps, plus cost implications for open source and base models.

Trace the evolution from neural networks to recurrent neural networks, showing how memory and feedback loops enable time-aware, sequence modeling, with LSTM capturing long-term dependencies.

Explore recurrent neural networks, including one-to-one, one-to-many, many-to-one, and many-to-many tasks, the encoder–decoder architecture, and how LSTMs address vanishing and exploding gradients.

Transformers address RNN and LSTM drawbacks by using attention to weigh relationships among all words, enabling parallelizable, contextual learning that powers large language models.

Learn how tokenizers map text to numerical values for transformers, compare word, subword, and character-based methods, and review byte pair encoding, SentencePiece, and BERT tokenizer with Hugging Face.

Explore embeddings and tokenization to encode word meaning and context, then examine encoder-decoder architecture with multi-head attention, encoder and decoder steps, and softmax predictions.

Explore encoder-only, decoder-only, and encoder-decoder transformers, compare bidirectional and unidirectional training objectives, and learn how to select models for sentiment analysis, translation, and summarization.

Explore decoding techniques for decoder-only transformers, including greedy search, beam search, and sampling methods like top K and top P, with practical GPT-2 examples in PyTorch.

Adapt and align models to your business using prompt engineering, fine-tuning, and human feedback. Explore data quality layers and prompt techniques like context, instructions, tasks, and few-shot.

Explore how autoencoders compress input data into a latent code and reconstruct it with an encoder and decoder, highlighting nonlinear dimensionality reduction and anomaly detection applications.

Build a Streamlit chatbot powered by LangChain and GPT to query a local SQLite database from the Looker ecommerce dataset, generating SQL and plotting results.