Mastering LLM Data Preparation with LLM DataStudio: A Step-by-Step Guide
- Generative AI
- H2O Cascade
Overview
H2O.ai's Head of Global Training, Andreea Turcu, walks through why data preparation is often the hardest part of working with large language models and how LLM Data Studio addresses it. The session covers what makes a model "large," the difference between LLMs, generative AI, and foundation models, and the LLM build lifecycle. Andreea then demos the tool's Curate, Prepare, and Augment tabs, including a live pipeline that turns a raw PDF into hundreds of cleaned question-answer pairs ready for fine-tuning.
Key takeaways
- The difference between large language models, generative AI, and foundation models, and why "large" refers to a model's number of parameters
- Why data preprocessing typically consumes around 80% of a data scientist's time in the LLM life cycle, and why clean input data drives model quality
- How to use LLM Data Studio's Curate tab to convert raw PDFs, audio, video, or web pages into question-answer datasets
- How the Prepare tab's Workflow Builder lets you drag and drop data-cleaning steps like profanity checks, length checks, and quality checks, and compare dataset quality before and after
- How the Augment tab enriches a dataset with additional or reinforcement-learning-style data to reduce bias
- A live walkthrough converting a 41-page PDF into 644 cleaned question-answer pairs, including a relevance score for each pair
Start owning your AI