LLMs with Google Cloud and Python

Set up your Google Cloud account through the Cloud Console and access 300 free credits over 90 days while learning about projects, billing, Cloud Shell, and Vertex AI API later.

Explore Vertex AI's text, code, and embedding models, palm, palm two, bison, and gecko. Learn pricing per 1000 characters and the foundation model naming scheme to choose the right option.

Learn zero shot, one shot, and few shot prompting to shape model responses. Apply prompt engineering, structured prompts, and input-output examples, using Google Cloud Console and Python API contexts.

Explore how the temperature parameter controls randomness in text generation, from deterministic outputs at zero to creative results at higher values, including its relation to top p and top k.

Explore how top-k and top-p filters shape the next-token distribution, then adjust temperature to balance determinism and creativity in generated text.

Explore stop sequences to hard stop decoding with custom strings, using practical prompts to stop at a number or a specific token, and combine with max tokens for precise outputs.

Explore one- and few-shot prompts for text generation and chat models using API calls, including one-shot templates, input-output pairs, and stop sequences.

Build a customer service chatbot with Vertex AI language models and Python, using a return policy context (30 days, unused) and a simple while loop to end the conversation.

Master vertex code APIs for code generation, chat, and completion. Match them to code bison, code chat bison, and code gecko across the software development life cycle.

Explore code prompts with Kodi, a Palm API based code generation model that writes functions, tests, and docstrings in languages like Python, Java, and JavaScript. Customize prompts for reliable code.

Explore code chat workflows using the code chat bison model to enable multi-turn conversations for writing, updating functions, and explaining code in notebooks.

Interact with a penguin CSV dataset by translating natural language queries into SQL using Vertex AI, pandas, and SQLAlchemy to query an in-memory database.

Master prompt engineering to elicit quality outputs from large language models by crafting inputs with context and examples, using input types like question, task, entity, and completion to mitigate hallucinations.

Master prompt structure by separating the task input from context or examples using newlines and indentation, and apply one-shot or few-shot formats to shape outputs in LLMs.

Learn how large language models excel at summarization by structuring prompts with the task and context, and tailor outputs as three sentences or five bullet points.

Explore how large language models classify text beyond sentiment analysis, using tweets about securities to assign flexible categories like finance, housing, and sports.

Learn to perform extraction tasks by combining classification, categorization, and prompt engineering to extract two stock tickers and their sentiments in an exact format.

Explore ideation and expansion with large language models by generating ideas, creating outlines, and expanding topics into blog posts, using prompts, token limits, and chat models for dynamic back-and-forth.

Explore transforming text with large language models by changing tone and format, from paragraphs to bullet lists, with hands-on examples and notes on model hallucinations.

Understand model hallucinations and distinguish plausible but incorrect outputs from truth, and use prompt engineering, probabilistic certainty checks, and fallback positions with context injection to avoid uncertain answers.

Translate a Spanish customer support email to English, then generate a concise summary and suggested next steps using a large language model. It demonstrates prompt engineering in a notebook.

Learn how text embeddings turn text into vector representations, enabling cosine similarity search, clustering, recommendations, anomaly detection, and injecting context from documents into prompts.

Inject context from a companies.csv into a large language model prompt using text embeddings and a summary function. Use cosine similarity for automatic context retrieval to reduce hallucinations.

Create and use text embeddings with a stable model version to enable reliable vector similarity searches, embedding summaries and linking them to their original texts for efficient retrieval.

Perform a similarity search using dot product on normalized embeddings to get cosine similarity, then inject the most similar embedding’s text as context to answer.

Discover how custom tuning tailors foundation models to specialized tasks by fine-tuning on task-specific data. Build JSON lines datasets for Vertex AI tuning and compare supervised tuning with RLHF.

Apply a practical custom tuning example using a json lines dataset in a Google Cloud Storage bucket. Configure regions, supervised tuning, and training steps to deploy a tuned model.