Fine-tuned large language models (LLMs) excel at relatively simple code generation tasks, but struggle with more intricate code, specialized libraries or complex application demos. This session addresses a well-known limitation of LLM—hallucinations—by proposing an innovative dataset representation strategy. While conventional fine-tuning often employs a question-output pair format, it also sometimes leads to undesirable hallucinations.
To combat these limitations, this session advocates for breaking down the training data into smaller, more descriptive components and constructing them sequentially into complex code. By adopting this structured approach, even smaller LLMs compatible with consumer-grade GPUs can improve significantly. Before fine-tuning, LLMs tend to generate random permutations of parameters and pipelines unrelated to the target library. However, after fine-tuning with an enhanced dataset representation, the model outputs code more aligned with the intended library, eliminating random permutations of library parameters from the output.
The session will discuss real-world experiences fine-tuning existing code generators to output high-caliber code on newer generative AI libraries such as Langchain, Vertex AI and others. It will offer a use case on the importance of data representation, as well as a practice strategy for enhancing LLM performance in code generation—particularly for complex and specialized tasks.
3 audience takeaways for this session:
- Improve LLM code generation accuracy
- Enhance LLM performance
- Understand real-world enterprise use cases for improved code generation
