DeepLearning.AI short courses

Improving the accuracy of LLM applications

Practical examples in this course were based on using Llama 3 with the Lamini library.

• Start with rigorous evaluation and iterating the prompt

• If adjusting the prompt isn't enough then try fine-tuning

• Fine-tuning often doesn't require much data

• Parameter-efficient fine-tuning can be very cheap

• Memory tuning allows you to embed specific facts directly into the model

  • I haven't independently researched whether this is actually useful or just Lamini marketing

• Evaluation dataset

  • Start small

  • Quality > Quantity

  • Focus on the areas that it does poorly on

  • Try to find the easiest examples that still fail

  • Try to break the process to identify issues

  • Set an accuracy target

  • Iterate the dataset as performance improves

• Scoring using LLMs

  • You would ideally use a deterministic approach instead if practical

  • Ask for a numerical score

  • Use a structured output to enforce this

  • Can you provide a reference answer for it to score against?

• Rolling your own fine-tuning can be hard, so consider managed fine-tuning

  • Inefficient implementations

  • Idle compute due to not being able to parallelize efficiently

  • Crashes

• Consider using LLMs to help create fine-tuning datasets