LLM Customization
summary
Compares concrete AWS paths: Bedrock vs SageMaker and Nova Forge, lists techniques used (SFT, preference alignment, RL fine-tuning, CPT, PEFT/LoRA) and gives data and time scales — SFT 10–10M examples, preference 1K–1M, CPT up to 1B tokens — states typical training runs take days to months, shows SageMaker Studio and serverless model customization for managed training, warns about catastrophic forgetting, infrastructure and versioning burdens, and reports that smaller customized models often cut latency while full-weight updates are costlier.
AI-written summary · about
chapters (7)
- 0:00The case for customization
- 2:24The customization spectrum
- 5:58Fine-tuning techniques
- 11:45Investment and strategy
- 16:33SageMaker infrastructure
- 22:14Implementation on SageMaker
- 26:39Nova Forge and wrap-up
related talks
37mStreamlining Machine Learning Operations
34mLat Designing and Building Complex Machine Learning....
33mBuild, train, and deploy ML models using Amazon SageMaker
31mTrain and tune ML models to the highest accuracy using Amazon SageMaker
30mLLMs: Beyond the Hype - A Practical Journey to Scale
39mTrain ML models at scale with Amazon SageMaker Training
33mMachine Learning Engineering Done Right
40mScaling generative AI workloads with efficient model choice