Fine-Tuning and Deploying LLMs with H2O Tools
- Generative AI
- Enterprise h2oGPTe
- H2O MLOps
- H2O Cascade
This lesson walks through the full workflow for preparing data, fine-tuning, and deploying a large language model with H2O tools, presented by H2O's Principal Customer Data Scientist Audrey. It covers when fine-tuning is actually needed versus using prompt engineering or a RAG system with Enterprise h2oGPTe, then demonstrates capturing preferred answers as training data, cleaning and preparing that data, and using causal language modeling with LoRA in H2O Cascade to fine-tune a Llama 2 based model. The lesson closes by pushing the tuned model to Hugging Face and registering and deploying it through H2O MLOps for API based scoring.
Key takeaways
- Why memorizing facts is a poor reason to fine-tune a model, since training data can quickly become outdated
- How to turn RAG feedback from Enterprise h2oGPTe into paired question and answer training data for fine-tuning
- What LoRA (Low-Rank Adaptation) does during fine-tuning to reduce memory footprint and speed up training in H2O Cascade
- When a hybrid approach combining RAG grounding and a fine-tuned writing style fits a use case better than either alone
- The steps for registering a fine-tuned model from Hugging Face into H2O MLOps and deploying it as a scoring endpoint
Start owning your AI