Abstract
State-of-the-art open models are very capable out of the box, but you can get even better results by adapting them to your domain, data, and use cases.
This talk covers widely used post-training techniques including supervised fine-tuning (SFT), LoRA, distillation, and related approaches. We’ll look at when to use each technique, the tradeoffs involved, and practical considerations for improving model quality without training a model from scratch.
Who it’s for
AI engineers, ML engineers, and developers who want to adapt open models for their own applications and domains.
What you’ll learn
- Understand why and when to post-train an open model.
- Compare common techniques including SFT, LoRA, and distillation.
- Understand the data, compute, and model requirements for each approach.
- Evaluate the tradeoffs between model quality, training cost, inference cost, and complexity.
- Choose an appropriate post-training strategy for your use case.
Format / duration
- Format: Talk + demo.
- Duration: 30 mins to 45 mins, adaptable to event format