Apr 16
Loading video player...
W&B Training offers Serverless SFT powered by CoreWeave to help AI engineers fine-tune large language models for agentic tasks without managing infrastructure. In this video, we show how Serverless SFT makes it faster to customize model output format and style, distill knowledge from curated datasets, and warm-start models for reinforcement learning in a unified post-training workflow. We also demonstrate how fine-tuned LoRA adapters can be served using W&B Inference for evaluation and deployment. *https://wandb.ai/site/serverless-sft* ⏳Timestamps: 0:00 Introducing W&B Training Serverless SFT powered by CoreWeave 0:25 AI applications are hard to productionize 1:23 Post-training LLMs with SFT and RL 2:16 Why switching between SFT and RL is difficult 2:46 Using SFT and RL in a unified workflow with W&B Training 3:51 Simple coding agent example 4:39 Evaluating coding agent LLMs 5:32 Getting started with Serverless SFT 6:09 Fine-tuning a Qwen model using Serverless SFT 7:36 Running Weave Evaluations during SFT 8:33 Post-training using SFT and RL together 9:32 Serving fine-tuned models using W&B Inference 9:55 Testing our fine-tuned model in the Weave Playground 10:27 Recap, conclusion, and invitation to try the Weights & Biases AI developer platform