Mastering Classification & Regression with LLMs: Insights from Kaggle Competitions | Philipp Singer
- Generative AI
Overview
In this session, H2O.ai's Philipp Singer explores how large language models can be applied to classical classification and regression problems, not just generative use cases. Using a public financial sentiment dataset from Kaggle, he compares a majority-class baseline, zero-shot LLM prompting, a next-token-probability trick for faster inference, and full fine-tuning with H2O LLM Studio. He also shares results from Kaggle competitions, including a first-place finish in the LLM Science Exam competition, where LLM-based classification outperformed traditional methods.
Key takeaways
- Why LLMs, originally built for generative tasks, can also be applied to classical text classification and regression problems
- How to get quick zero-shot predictions from an LLM, and the prompt-engineering tradeoffs that make it hard to productionize
- A next-token-probability trick that turns a zero-shot LLM into a fast classifier without parsing free-text output
- How to fine-tune LLMs for classification and regression in H2O LLM Studio, including adding a custom classification head
- Benchmark results from H2O LLM Studio experiments, reaching roughly 94 to 95 percent accuracy on a sentiment task in minutes
- How LLM-based classification performed in Kaggle competitions, including a first-place finish in the LLM Science Exam competition
Start owning your AI