10 key takeaways from the session👇
1. 🤔 ~ 50% people have the most trouble explaining #instruction #tuning vs. #finetuning and #promptengineering
2. 🤯 #instruction #tuning IS #finetuning, it is simply a subset of all possible fine-tuning.
3. 🎛 When we discuss #finetuning of #llms, we're talking about changing the input-output schema
4. 👾 The best visual for understanding any supervised #finetuning of an #llm trained by #unsupervised pre-training is Shoggoth (en.wikipedia.org/wiki/Shoggoth)
5. 🥇 The most important #promptengineering best practices are 1. Instruction and 2. Role (especially if you're working as a developer using tools like LangChain, where role is fundamental)
6. ♻ Beyond instruction and role, focus on your ability to work with #llms to do self-refinement. Self-refinement will help you move beyond zero-shot learning and even beyond few-shot learning to chain-of-thought reasoning in your prompts.
Self-Refinement Example: colab.research.google.com/dri...
⛓ Chain-of-Thought Prompting Example: colab.research.google.com/dri...
7. 🕳 Prompting is best for getting solutions up and running quickly, and is a way to investigate the latent spaces within Shoggoth (the #llm).
8. 📈 Use #finetuning of I/O schema when you're nearing the maximum context window for prompts, once you've determined the #userexperience and interaction that you want users to have with the LLM, or when you want to use many examples to give better model performance.
9. 1️⃣ The no. 1 rule in #promptengineering is Instruction, and if you're building with an #LLM you should always pick up an Instruction-tuned one over a base model. This is not a coincidence.
10. 🌐 One of the most underrated ways to improve prompts is to use web-search-enabled GPT models like Bing - have them read prompting guides and then suggest specific improvements.
BONUS DEMO!
- Check out how to Instruction-Tune Lit-LLaMA on the wikisql dataset to produce SQL code from Natural Language, by Chris Alexiuk at the end of the session.
- Colab Link: colab.research.google.com/dri...
Slide Deck: docs.google.com/presentation/...
Негізгі бет Ғылым және технология Community Series: Generative AI and Large Language Models: Prompt Engineering vs. Fine-Tuning
Пікірлер