Make Your Own GPT with h2oGPT and H2O LLM Studio
- Generative AI
- H2O Cascade
This lesson recaps a webinar where H2O data scientist Rob Mulla walks through building a custom large language model with h2oGPT and H2O LLM Studio. He covers LLM fundamentals, including model size and training data, then breaks customization into three levels: prompt engineering, grounding with retrieval, and fine tuning. The session includes a live demo of h2oGPT answering grounded questions from an uploaded document and a walkthrough of fine tuning a model in H2O LLM Studio using LoRA. It closes with guidance on picking the right customization approach for the task at hand.
Key takeaways
- How the three levels of LLM customization differ: prompt engineering, grounding with retrieval, and fine tuning
- Why larger foundational models are not automatically better, since training data and task objectives matter as much as parameter count
- How h2oGPT uses a vector database to ground answers in uploaded documents and cite its sources
- What H2O LLM Studio does: fine tunes open source models with LoRA and lets you chat with the model while it trains
- The difference between in context grounding, which adds new information, and fine tuning, which changes tone and response style
- How quantization lets larger models run on smaller GPUs by reducing the precision of stored weights
Start owning your AI