Pimp My Llama - How to Customize LLMs for your Business Case
summary
A live Jonas Simulator Q&A is built from a vanilla Llama 3 through three concrete steps: tight prompt templates that stop obvious errors, a RAG pipeline retrieving about 3,000 documents to reduce hallucinations, and optional fine‑tuning (including parameter‑efficient LoRA instead of full retrain) for domain style. Hardware needs and limits appear: models with ~70B parameters force multi‑GPU inference and quantization tradeoffs; the demo runs and answers factual queries but still misses some aggregated meta‑counts.
AI-written summary · about
chapters (8)
- 0:00Pimping LLMs for business
- 1:49Innovation hacking at TNG
- 4:36Understanding LLMs
- 13:14Prompt engineering
- 22:19RAG for factual context
- 33:35Fine-tuning LLMs
- 39:50Conclusion and demo
- 40:33Q&A
related talks
50mPimp My Llama - How to Customize LLMs for Your Business Case
49mTame your Llama, run AI in Java
09mTailoring Llama for Your Needs: Fine-Tuning & Evaluation
48mBringing Llama 3 to Life
12mUnpack the Latest Llama 4 Herd
55mBuild enterprise generative AI apps using Llama 3 at 1,000 tokens/s on the SambaNova AI platform
48mPracticalities of Fine-Tuning Llama 2 with AI Studio
29mLocal LLM fine-tuning: a practical example