https://www.coursera.org/learn/generative-ai-with-llms/supplement/b1kYM/lecture-notes-week-2

AI Discussions
1

Hi everyone! :waving_hand:
I’m La Mema Parandy, and I’m exploring the Week 2 labs on fine-tuning LLMs for dialogue summarization. I’ve been learning about full fine-tuning, PEFT, LoRA, and soft prompt tuning, and I’m really curious about how others are applying these techniques.

Here are a few questions I’d love to discuss with you all:

  1. Fine-Tuning Trade-offs: How do full fine-tuning and PEFT compare in your projects in terms of performance vs. computational resources?

  2. Instruction & Multi-Task Fine-Tuning: How could these approaches be applied to real-world NLP tools like chatbots or summarization systems?

  3. Evaluation Metrics: Which metrics (ROUGE, BLEU, HELM, GLUE/SuperGLUE) do you rely on most for judging summarization quality?

  4. LoRA & Soft Prompt Tuning: Has anyone tried these methods yet? How did the performance improvements compare to resource savings?

  5. Applications & Ethics: How can we apply fine-tuned models responsibly while minimizing bias?

I’d love to hear your experiences, thoughts, or even resources you’ve found helpful. Looking forward to learning from you all!

2

Dear @papalame99,

Welcome to the Community!

We’re excited to have you here.

  • Full Fine-Tuning: Delivers the best performance, but it’s very resource-intensive and expensive.
  • PEFT/LoRA: Offers around 95% of full fine-tuning performance at a much lower cost, making it perfect for most real-world tasks.
  • Instruction Tuning: Works really well for chatbots and summarizers, especially when paired with RLHF for better alignment.
  • Metrics: Use ROUGE + human feedback for summarization, GLUE for general NLP tasks, and HELM to check for bias and safety.
  • LoRA vs. Soft Prompts: LoRA gives stronger results, while Soft Prompts are great for quick, lightweight customization.
  • Ethics: Always use diverse datasets, actively check for bias, and put safety measures in place before deployment.

In short: LoRA is a great balance of performance and efficiency, while full fine-tuning is best for highly specialized, critical domains.

Looking forward to learning and growing together.

–
Keep Learning AI with DeepLearning.AI - Girijesh