Post-Training on CoreWeave Forge

David Corbitt introduces CoreWeave Forge post-training, explaining how his background co-founding OpenPipe led to CoreWeave acquiring the company to bring accessible fine-tuning to enterprises. Forge provides an integrated environment where observability, post-training, and inference work together in a continuous improvement cycle. Teams observe production behavior, extract traces to form evaluation or training datasets, alter the weights of open-source models, and redeploy. The goal of the managed platform is to eliminate the need for in-house machine learning experts or deep PyTorch knowledge by handling the underlying complexities of training runs and quickly catching errors through integrated evaluations and metrics.

The presentation covers three supported post-training methods: supervised fine-tuning, reinforcement learning, and model distillation. Supervised fine-tuning relies on high-quality input and output examples to enforce style or consistency, as seen with startup Willow using it to clean up voice-to-text transcriptions containing self-corrections. Reinforcement learning uses reward functions to improve model performance, evidenced by a case study with You.com and Nvidia that increased web search accuracy while reducing tool calls. Corbitt emphasizes that before attempting post-training, users should always consider faster and cheaper alternatives, such as optimizing prompts, repairing environment tools, or evaluating larger models.

CoreWeave manages this process on its own GPU infrastructure, where base models remain pre-loaded to avoid spin-up latency and out-of-memory issues. Instead of retraining full base weights, the platform trains small LoRA adapters that run alongside shared models on serverless inference, allowing customers to pay standard token rates for customized performance. For distillation, an automated background agent gathers production data, relabels it with a stronger model, conducts training sweeps on a smaller model, and evaluates the results against production standards. This architecture establishes an automated, repeatable lifecycle that reduces deployment costs and operational complexity while allowing models to adapt to changes over time.