Keep pulling the thread on WNB Training Serverless SFT.
WNB Training Serverless SFT allows AI engineers to alternate between SFT and RL training without moving model artifacts across different systems.
WNB Training allows users to start serverless RL runs from optimal SFT checkpoints.
WNB Training Serverless SFT is powered by CoreWeave.
WNB Training Serverless SFT is designed to help AI engineers with model distillation, customizing model output format, and preparing models for reinforcement learning training.
WNB Training provides engineers with instant access to CoreWeave GPU capacity, with provisioning and scaling handled automatically.
Users can initiate a training run on the WNB platform by calling the open-source Agent Reinforcement Trainer (ART) API.
During and after a Supervised Fine-Tuning run, the resulting LoRa adapters are saved directly to WNB Artifacts.
In an internal evaluation of a coding agent, a QIN3 base model demonstrated lower accuracy than GPT models.
In an internal evaluation of a coding agent, a QIN3 base model had better latency and cost performance compared to GPT models.
The WNB platform allows for running Weave evaluations on fine-tuned LoRa weights after every training epoch.
Users can embed Weave evaluation panels directly into their WNB workspace to visualize results.
Production-candidate LoRa weights can be deployed for final testing and production using WNB Inference.